1 /*-------------------------------------------------------------------------
5 * PostgreSQL Integrated Autovacuum Daemon
7 * The autovacuum system is structured in two different kinds of processes: the
8 * autovacuum launcher and the autovacuum worker. The launcher is an
9 * always-running process, started by the postmaster when the autovacuum GUC
10 * parameter is set. The launcher schedules autovacuum workers to be started
11 * when appropriate. The workers are the processes which execute the actual
12 * vacuuming; they connect to a database as determined in the launcher, and
13 * once connected they examine the catalogs to select the tables to vacuum.
15 * The autovacuum launcher cannot start the worker processes by itself,
16 * because doing so would cause robustness issues (namely, failure to shut
17 * them down on exceptional conditions, and also, since the launcher is
18 * connected to shared memory and is thus subject to corruption there, it is
19 * not as robust as the postmaster). So it leaves that task to the postmaster.
21 * There is an autovacuum shared memory area, where the launcher stores
22 * information about the database it wants vacuumed. When it wants a new
23 * worker to start, it sets a flag in shared memory and sends a signal to the
24 * postmaster. Then postmaster knows nothing more than it must start a worker;
25 * so it forks a new child, which turns into a worker. This new process
26 * connects to shared memory, and there it can inspect the information that the
27 * launcher has set up.
29 * If the fork() call fails in the postmaster, it sets a flag in the shared
30 * memory area, and sends a signal to the launcher. The launcher, upon
31 * noticing the flag, can try starting the worker again by resending the
32 * signal. Note that the failure can only be transient (fork failure due to
33 * high load, memory pressure, too many processes, etc); more permanent
34 * problems, like failure to connect to a database, are detected later in the
35 * worker and dealt with just by having the worker exit normally. The launcher
36 * will launch a new worker again later, per schedule.
38 * When the worker is done vacuuming it sends SIGUSR2 to the launcher. The
39 * launcher then wakes up and is able to launch another worker, if the schedule
40 * is so tight that a new worker is needed immediately. At this time the
41 * launcher can also balance the settings for the various remaining workers'
42 * cost-based vacuum delay feature.
44 * Note that there can be more than one worker in a database concurrently.
45 * They will store the table they are currently vacuuming in shared memory, so
46 * that other workers avoid being blocked waiting for the vacuum lock for that
47 * table. They will also fetch the last time the table was vacuumed from
48 * pgstats just before vacuuming each table, to avoid vacuuming a table that
49 * was just finished being vacuumed by another worker and thus is no longer
50 * noted in shared memory. However, there is a small window (due to not yet
51 * holding the relation lock) during which a worker may choose a table that was
52 * already vacuumed; this is a bug in the current design.
54 * Portions Copyright (c) 1996-2023, PostgreSQL Global Development Group
55 * Portions Copyright (c) 1994, Regents of the University of California
59 * src/backend/postmaster/autovacuum.c
61 *-------------------------------------------------------------------------
69 #include "access/heapam.h"
70 #include "access/htup_details.h"
71 #include "access/multixact.h"
72 #include "access/reloptions.h"
73 #include "access/tableam.h"
74 #include "access/transam.h"
75 #include "access/xact.h"
76 #include "catalog/dependency.h"
77 #include "catalog/namespace.h"
78 #include "catalog/pg_database.h"
79 #include "commands/dbcommands.h"
80 #include "commands/vacuum.h"
81 #include "lib/ilist.h"
82 #include "libpq/pqsignal.h"
83 #include "miscadmin.h"
84 #include "nodes/makefuncs.h"
86 #include "postmaster/autovacuum.h"
87 #include "postmaster/fork_process.h"
88 #include "postmaster/interrupt.h"
89 #include "postmaster/postmaster.h"
90 #include "storage/bufmgr.h"
91 #include "storage/ipc.h"
92 #include "storage/latch.h"
93 #include "storage/lmgr.h"
94 #include "storage/pmsignal.h"
95 #include "storage/proc.h"
96 #include "storage/procsignal.h"
97 #include "storage/sinvaladt.h"
98 #include "storage/smgr.h"
99 #include "tcop/tcopprot.h"
100 #include "utils/fmgroids.h"
101 #include "utils/fmgrprotos.h"
102 #include "utils/guc_hooks.h"
103 #include "utils/lsyscache.h"
104 #include "utils/memutils.h"
105 #include "utils/ps_status.h"
106 #include "utils/rel.h"
107 #include "utils/snapmgr.h"
108 #include "utils/syscache.h"
109 #include "utils/timeout.h"
110 #include "utils/timestamp.h"
116 bool autovacuum_start_daemon
= false;
117 int autovacuum_max_workers
;
118 int autovacuum_work_mem
= -1;
119 int autovacuum_naptime
;
120 int autovacuum_vac_thresh
;
121 double autovacuum_vac_scale
;
122 int autovacuum_vac_ins_thresh
;
123 double autovacuum_vac_ins_scale
;
124 int autovacuum_anl_thresh
;
125 double autovacuum_anl_scale
;
126 int autovacuum_freeze_max_age
;
127 int autovacuum_multixact_freeze_max_age
;
129 double autovacuum_vac_cost_delay
;
130 int autovacuum_vac_cost_limit
;
132 int Log_autovacuum_min_duration
= 600000;
134 /* the minimum allowed time between two awakenings of the launcher */
135 #define MIN_AUTOVAC_SLEEPTIME 100.0 /* milliseconds */
136 #define MAX_AUTOVAC_SLEEPTIME 300 /* seconds */
138 /* Flags to tell if we are in an autovacuum process */
139 static bool am_autovacuum_launcher
= false;
140 static bool am_autovacuum_worker
= false;
142 /* Flags set by signal handlers */
143 static volatile sig_atomic_t got_SIGUSR2
= false;
145 /* Comparison points for determining whether freeze_max_age is exceeded */
146 static TransactionId recentXid
;
147 static MultiXactId recentMulti
;
149 /* Default freeze ages to use for autovacuum (varies by database) */
150 static int default_freeze_min_age
;
151 static int default_freeze_table_age
;
152 static int default_multixact_freeze_min_age
;
153 static int default_multixact_freeze_table_age
;
155 /* Memory context for long-lived data */
156 static MemoryContext AutovacMemCxt
;
158 /* struct to keep track of databases in launcher */
159 typedef struct avl_dbase
161 Oid adl_datid
; /* hash key -- must be first */
162 TimestampTz adl_next_worker
;
167 /* struct to keep track of databases in worker */
168 typedef struct avw_dbase
172 TransactionId adw_frozenxid
;
173 MultiXactId adw_minmulti
;
174 PgStat_StatDBEntry
*adw_entry
;
177 /* struct to keep track of tables to vacuum and/or analyze, in 1st pass */
178 typedef struct av_relation
180 Oid ar_toastrelid
; /* hash key - must be first */
183 AutoVacOpts ar_reloptions
; /* copy of AutoVacOpts from the main table's
184 * reloptions, or NULL if none */
187 /* struct to keep track of tables to vacuum and/or analyze, after rechecking */
188 typedef struct autovac_table
191 VacuumParams at_params
;
192 double at_vacuum_cost_delay
;
193 int at_vacuum_cost_limit
;
202 * This struct holds information about a single worker's whereabouts. We keep
203 * an array of these in shared memory, sized according to
204 * autovacuum_max_workers.
206 * wi_links entry into free list or running list
207 * wi_dboid OID of the database this worker is supposed to work on
208 * wi_tableoid OID of the table currently being vacuumed, if any
209 * wi_sharedrel flag indicating whether table is marked relisshared
210 * wi_proc pointer to PGPROC of the running worker, NULL if not started
211 * wi_launchtime Time at which this worker was launched
212 * wi_cost_* Vacuum cost-based delay parameters current in this worker
214 * All fields are protected by AutovacuumLock, except for wi_tableoid and
215 * wi_sharedrel which are protected by AutovacuumScheduleLock (note these
216 * two fields are read-only for everyone except that worker itself).
219 typedef struct WorkerInfoData
225 TimestampTz wi_launchtime
;
228 double wi_cost_delay
;
230 int wi_cost_limit_base
;
233 typedef struct WorkerInfoData
*WorkerInfo
;
236 * Possible signals received by the launcher from remote processes. These are
237 * stored atomically in shared memory so that other processes can set them
242 AutoVacForkFailed
, /* failed trying to start a worker */
243 AutoVacRebalance
, /* rebalance the cost limits */
244 AutoVacNumSignals
/* must be last */
248 * Autovacuum workitem array, stored in AutoVacuumShmem->av_workItems. This
249 * list is mostly protected by AutovacuumLock, except that if an item is
250 * marked 'active' other processes must not modify the work-identifying
253 typedef struct AutoVacuumWorkItem
255 AutoVacuumWorkItemType avw_type
;
256 bool avw_used
; /* below data is valid */
257 bool avw_active
; /* being processed */
260 BlockNumber avw_blockNumber
;
261 } AutoVacuumWorkItem
;
263 #define NUM_WORKITEMS 256
266 * The main autovacuum shmem struct. On shared memory we store this main
267 * struct and the array of WorkerInfo structs. This struct keeps:
269 * av_signal set by other processes to indicate various conditions
270 * av_launcherpid the PID of the autovacuum launcher
271 * av_freeWorkers the WorkerInfo freelist
272 * av_runningWorkers the WorkerInfo non-free queue
273 * av_startingWorker pointer to WorkerInfo currently being started (cleared by
274 * the worker itself as soon as it's up and running)
275 * av_workItems work item array
277 * This struct is protected by AutovacuumLock, except for av_signal and parts
278 * of the worker list (see above).
283 sig_atomic_t av_signal
[AutoVacNumSignals
];
284 pid_t av_launcherpid
;
285 dlist_head av_freeWorkers
;
286 dlist_head av_runningWorkers
;
287 WorkerInfo av_startingWorker
;
288 AutoVacuumWorkItem av_workItems
[NUM_WORKITEMS
];
289 } AutoVacuumShmemStruct
;
291 static AutoVacuumShmemStruct
*AutoVacuumShmem
;
294 * the database list (of avl_dbase elements) in the launcher, and the context
297 static dlist_head DatabaseList
= DLIST_STATIC_INIT(DatabaseList
);
298 static MemoryContext DatabaseListCxt
= NULL
;
300 /* Pointer to my own WorkerInfo, valid on each worker */
301 static WorkerInfo MyWorkerInfo
= NULL
;
303 /* PID of launcher, valid only in worker while shutting down */
304 int AutovacuumLauncherPid
= 0;
307 static pid_t
avlauncher_forkexec(void);
308 static pid_t
avworker_forkexec(void);
310 NON_EXEC_STATIC
void AutoVacWorkerMain(int argc
, char *argv
[]) pg_attribute_noreturn();
311 NON_EXEC_STATIC
void AutoVacLauncherMain(int argc
, char *argv
[]) pg_attribute_noreturn();
313 static Oid
do_start_worker(void);
314 static void HandleAutoVacLauncherInterrupts(void);
315 static void AutoVacLauncherShutdown(void) pg_attribute_noreturn();
316 static void launcher_determine_sleep(bool canlaunch
, bool recursing
,
317 struct timeval
*nap
);
318 static void launch_worker(TimestampTz now
);
319 static List
*get_database_list(void);
320 static void rebuild_database_list(Oid newdb
);
321 static int db_comparator(const void *a
, const void *b
);
322 static void autovac_balance_cost(void);
324 static void do_autovacuum(void);
325 static void FreeWorkerInfo(int code
, Datum arg
);
327 static autovac_table
*table_recheck_autovac(Oid relid
, HTAB
*table_toast_map
,
328 TupleDesc pg_class_desc
,
329 int effective_multixact_freeze_max_age
);
330 static void recheck_relation_needs_vacanalyze(Oid relid
, AutoVacOpts
*avopts
,
331 Form_pg_class classForm
,
332 int effective_multixact_freeze_max_age
,
333 bool *dovacuum
, bool *doanalyze
, bool *wraparound
);
334 static void relation_needs_vacanalyze(Oid relid
, AutoVacOpts
*relopts
,
335 Form_pg_class classForm
,
336 PgStat_StatTabEntry
*tabentry
,
337 int effective_multixact_freeze_max_age
,
338 bool *dovacuum
, bool *doanalyze
, bool *wraparound
);
340 static void autovacuum_do_vac_analyze(autovac_table
*tab
,
341 BufferAccessStrategy bstrategy
);
342 static AutoVacOpts
*extract_autovac_opts(HeapTuple tup
,
343 TupleDesc pg_class_desc
);
344 static void perform_work_item(AutoVacuumWorkItem
*workitem
);
345 static void autovac_report_activity(autovac_table
*tab
);
346 static void autovac_report_workitem(AutoVacuumWorkItem
*workitem
,
347 const char *nspname
, const char *relname
);
348 static void avl_sigusr2_handler(SIGNAL_ARGS
);
352 /********************************************************************
353 * AUTOVACUUM LAUNCHER CODE
354 ********************************************************************/
358 * forkexec routine for the autovacuum launcher process.
360 * Format up the arglist, then fork and exec.
363 avlauncher_forkexec(void)
368 av
[ac
++] = "postgres";
369 av
[ac
++] = "--forkavlauncher";
370 av
[ac
++] = NULL
; /* filled in by postmaster_forkexec */
373 Assert(ac
< lengthof(av
));
375 return postmaster_forkexec(ac
, av
);
379 * We need this set from the outside, before InitProcess is called
382 AutovacuumLauncherIAm(void)
384 am_autovacuum_launcher
= true;
389 * Main entry point for autovacuum launcher process, to be called from the
393 StartAutoVacLauncher(void)
398 switch ((AutoVacPID
= avlauncher_forkexec()))
400 switch ((AutoVacPID
= fork_process()))
405 (errmsg("could not fork autovacuum launcher process: %m")));
410 /* in postmaster child ... */
411 InitPostmasterChild();
413 /* Close the postmaster's sockets */
414 ClosePostmasterPorts(false);
416 AutoVacLauncherMain(0, NULL
);
420 return (int) AutoVacPID
;
423 /* shouldn't get here */
428 * Main loop for the autovacuum launcher process.
431 AutoVacLauncherMain(int argc
, char *argv
[])
433 sigjmp_buf local_sigjmp_buf
;
435 am_autovacuum_launcher
= true;
437 MyBackendType
= B_AUTOVAC_LAUNCHER
;
438 init_ps_display(NULL
);
441 (errmsg_internal("autovacuum launcher started")));
444 pg_usleep(PostAuthDelay
* 1000000L);
446 SetProcessingMode(InitProcessing
);
449 * Set up signal handlers. We operate on databases much like a regular
450 * backend, so we use the same signal handling. See equivalent code in
453 pqsignal(SIGHUP
, SignalHandlerForConfigReload
);
454 pqsignal(SIGINT
, StatementCancelHandler
);
455 pqsignal(SIGTERM
, SignalHandlerForShutdownRequest
);
456 /* SIGQUIT handler was already set up by InitPostmasterChild */
458 InitializeTimeouts(); /* establishes SIGALRM handler */
460 pqsignal(SIGPIPE
, SIG_IGN
);
461 pqsignal(SIGUSR1
, procsignal_sigusr1_handler
);
462 pqsignal(SIGUSR2
, avl_sigusr2_handler
);
463 pqsignal(SIGFPE
, FloatExceptionHandler
);
464 pqsignal(SIGCHLD
, SIG_DFL
);
467 * Create a per-backend PGPROC struct in shared memory, except in the
468 * EXEC_BACKEND case where this was done in SubPostmasterMain. We must do
469 * this before we can use LWLocks (and in the EXEC_BACKEND case we already
470 * had to do some stuff with LWLocks).
476 /* Early initialization */
479 InitPostgres(NULL
, InvalidOid
, NULL
, InvalidOid
, false, false, NULL
);
481 SetProcessingMode(NormalProcessing
);
484 * Create a memory context that we will do all our work in. We do this so
485 * that we can reset the context during error recovery and thereby avoid
486 * possible memory leaks.
488 AutovacMemCxt
= AllocSetContextCreate(TopMemoryContext
,
489 "Autovacuum Launcher",
490 ALLOCSET_DEFAULT_SIZES
);
491 MemoryContextSwitchTo(AutovacMemCxt
);
494 * If an exception is encountered, processing resumes here.
496 * This code is a stripped down version of PostgresMain error recovery.
498 * Note that we use sigsetjmp(..., 1), so that the prevailing signal mask
499 * (to wit, BlockSig) will be restored when longjmp'ing to here. Thus,
500 * signals other than SIGQUIT will be blocked until we complete error
501 * recovery. It might seem that this policy makes the HOLD_INTERRUPTS()
502 * call redundant, but it is not since InterruptPending might be set
505 if (sigsetjmp(local_sigjmp_buf
, 1) != 0)
507 /* since not using PG_TRY, must reset error stack by hand */
508 error_context_stack
= NULL
;
510 /* Prevents interrupts while cleaning up */
513 /* Forget any pending QueryCancel or timeout request */
514 disable_all_timeouts(false);
515 QueryCancelPending
= false; /* second to avoid race condition */
517 /* Report the error to the server log */
520 /* Abort the current transaction in order to recover */
521 AbortCurrentTransaction();
524 * Release any other resources, for the case where we were not in a
528 pgstat_report_wait_end();
531 /* this is probably dead code, but let's be safe: */
532 if (AuxProcessResourceOwner
)
533 ReleaseAuxProcessResources(false);
534 AtEOXact_Buffers(false);
536 AtEOXact_Files(false);
537 AtEOXact_HashTables(false);
540 * Now return to normal top-level context and clear ErrorContext for
543 MemoryContextSwitchTo(AutovacMemCxt
);
546 /* Flush any leaked data in the top-level context */
547 MemoryContextResetAndDeleteChildren(AutovacMemCxt
);
549 /* don't leave dangling pointers to freed memory */
550 DatabaseListCxt
= NULL
;
551 dlist_init(&DatabaseList
);
553 /* Now we can allow interrupts again */
556 /* if in shutdown mode, no need for anything further; just go away */
557 if (ShutdownRequestPending
)
558 AutoVacLauncherShutdown();
561 * Sleep at least 1 second after any error. We don't want to be
562 * filling the error logs as fast as we can.
567 /* We can now handle ereport(ERROR) */
568 PG_exception_stack
= &local_sigjmp_buf
;
570 /* must unblock signals before calling rebuild_database_list */
571 sigprocmask(SIG_SETMASK
, &UnBlockSig
, NULL
);
574 * Set always-secure search path. Launcher doesn't connect to a database,
575 * so this has no effect.
577 SetConfigOption("search_path", "", PGC_SUSET
, PGC_S_OVERRIDE
);
580 * Force zero_damaged_pages OFF in the autovac process, even if it is set
581 * in postgresql.conf. We don't really want such a dangerous option being
582 * applied non-interactively.
584 SetConfigOption("zero_damaged_pages", "false", PGC_SUSET
, PGC_S_OVERRIDE
);
587 * Force settable timeouts off to avoid letting these settings prevent
588 * regular maintenance from being executed.
590 SetConfigOption("statement_timeout", "0", PGC_SUSET
, PGC_S_OVERRIDE
);
591 SetConfigOption("lock_timeout", "0", PGC_SUSET
, PGC_S_OVERRIDE
);
592 SetConfigOption("idle_in_transaction_session_timeout", "0",
593 PGC_SUSET
, PGC_S_OVERRIDE
);
596 * Force default_transaction_isolation to READ COMMITTED. We don't want
597 * to pay the overhead of serializable mode, nor add any risk of causing
598 * deadlocks or delaying other transactions.
600 SetConfigOption("default_transaction_isolation", "read committed",
601 PGC_SUSET
, PGC_S_OVERRIDE
);
604 * Even when system is configured to use a different fetch consistency,
605 * for autovac we always want fresh stats.
607 SetConfigOption("stats_fetch_consistency", "none", PGC_SUSET
, PGC_S_OVERRIDE
);
610 * In emergency mode, just start a worker (unless shutdown was requested)
613 if (!AutoVacuumingActive())
615 if (!ShutdownRequestPending
)
617 proc_exit(0); /* done */
620 AutoVacuumShmem
->av_launcherpid
= MyProcPid
;
623 * Create the initial database list. The invariant we want this list to
624 * keep is that it's ordered by decreasing next_time. As soon as an entry
625 * is updated to a higher time, it will be moved to the front (which is
626 * correct because the only operation is to add autovacuum_naptime to the
627 * entry, and time always increases).
629 rebuild_database_list(InvalidOid
);
631 /* loop until shutdown request */
632 while (!ShutdownRequestPending
)
635 TimestampTz current_time
= 0;
639 * This loop is a bit different from the normal use of WaitLatch,
640 * because we'd like to sleep before the first launch of a child
641 * process. So it's WaitLatch, then ResetLatch, then check for
642 * wakening conditions.
645 launcher_determine_sleep(!dlist_is_empty(&AutoVacuumShmem
->av_freeWorkers
),
649 * Wait until naptime expires or we get some type of signal (all the
650 * signal handlers will wake us by calling SetLatch).
652 (void) WaitLatch(MyLatch
,
653 WL_LATCH_SET
| WL_TIMEOUT
| WL_EXIT_ON_PM_DEATH
,
654 (nap
.tv_sec
* 1000L) + (nap
.tv_usec
/ 1000L),
655 WAIT_EVENT_AUTOVACUUM_MAIN
);
659 HandleAutoVacLauncherInterrupts();
662 * a worker finished, or postmaster signaled failure to start a worker
668 /* rebalance cost limits, if needed */
669 if (AutoVacuumShmem
->av_signal
[AutoVacRebalance
])
671 LWLockAcquire(AutovacuumLock
, LW_EXCLUSIVE
);
672 AutoVacuumShmem
->av_signal
[AutoVacRebalance
] = false;
673 autovac_balance_cost();
674 LWLockRelease(AutovacuumLock
);
677 if (AutoVacuumShmem
->av_signal
[AutoVacForkFailed
])
680 * If the postmaster failed to start a new worker, we sleep
681 * for a little while and resend the signal. The new worker's
682 * state is still in memory, so this is sufficient. After
683 * that, we restart the main loop.
685 * XXX should we put a limit to the number of times we retry?
686 * I don't think it makes much sense, because a future start
687 * of a worker will continue to fail in the same way.
689 AutoVacuumShmem
->av_signal
[AutoVacForkFailed
] = false;
690 pg_usleep(1000000L); /* 1s */
691 SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_WORKER
);
697 * There are some conditions that we need to check before trying to
698 * start a worker. First, we need to make sure that there is a worker
699 * slot available. Second, we need to make sure that no other worker
700 * failed while starting up.
703 current_time
= GetCurrentTimestamp();
704 LWLockAcquire(AutovacuumLock
, LW_SHARED
);
706 can_launch
= !dlist_is_empty(&AutoVacuumShmem
->av_freeWorkers
);
708 if (AutoVacuumShmem
->av_startingWorker
!= NULL
)
711 WorkerInfo worker
= AutoVacuumShmem
->av_startingWorker
;
714 * We can't launch another worker when another one is still
715 * starting up (or failed while doing so), so just sleep for a bit
716 * more; that worker will wake us up again as soon as it's ready.
717 * We will only wait autovacuum_naptime seconds (up to a maximum
718 * of 60 seconds) for this to happen however. Note that failure
719 * to connect to a particular database is not a problem here,
720 * because the worker removes itself from the startingWorker
721 * pointer before trying to connect. Problems detected by the
722 * postmaster (like fork() failure) are also reported and handled
723 * differently. The only problems that may cause this code to
724 * fire are errors in the earlier sections of AutoVacWorkerMain,
725 * before the worker removes the WorkerInfo from the
726 * startingWorker pointer.
728 waittime
= Min(autovacuum_naptime
, 60) * 1000;
729 if (TimestampDifferenceExceeds(worker
->wi_launchtime
, current_time
,
732 LWLockRelease(AutovacuumLock
);
733 LWLockAcquire(AutovacuumLock
, LW_EXCLUSIVE
);
736 * No other process can put a worker in starting mode, so if
737 * startingWorker is still INVALID after exchanging our lock,
738 * we assume it's the same one we saw above (so we don't
739 * recheck the launch time).
741 if (AutoVacuumShmem
->av_startingWorker
!= NULL
)
743 worker
= AutoVacuumShmem
->av_startingWorker
;
744 worker
->wi_dboid
= InvalidOid
;
745 worker
->wi_tableoid
= InvalidOid
;
746 worker
->wi_sharedrel
= false;
747 worker
->wi_proc
= NULL
;
748 worker
->wi_launchtime
= 0;
749 dlist_push_head(&AutoVacuumShmem
->av_freeWorkers
,
751 AutoVacuumShmem
->av_startingWorker
= NULL
;
753 errmsg("autovacuum worker took too long to start; canceled"));
759 LWLockRelease(AutovacuumLock
); /* either shared or exclusive */
761 /* if we can't do anything, just go back to sleep */
765 /* We're OK to start a new worker */
767 if (dlist_is_empty(&DatabaseList
))
770 * Special case when the list is empty: start a worker right away.
771 * This covers the initial case, when no database is in pgstats
772 * (thus the list is empty). Note that the constraints in
773 * launcher_determine_sleep keep us from starting workers too
774 * quickly (at most once every autovacuum_naptime when the list is
777 launch_worker(current_time
);
782 * because rebuild_database_list constructs a list with most
783 * distant adl_next_worker first, we obtain our database from the
788 avdb
= dlist_tail_element(avl_dbase
, adl_node
, &DatabaseList
);
791 * launch a worker if next_worker is right now or it is in the
794 if (TimestampDifferenceExceeds(avdb
->adl_next_worker
,
796 launch_worker(current_time
);
800 AutoVacLauncherShutdown();
804 * Process any new interrupts.
807 HandleAutoVacLauncherInterrupts(void)
809 /* the normal shutdown case */
810 if (ShutdownRequestPending
)
811 AutoVacLauncherShutdown();
813 if (ConfigReloadPending
)
815 ConfigReloadPending
= false;
816 ProcessConfigFile(PGC_SIGHUP
);
818 /* shutdown requested in config file? */
819 if (!AutoVacuumingActive())
820 AutoVacLauncherShutdown();
822 /* rebalance in case the default cost parameters changed */
823 LWLockAcquire(AutovacuumLock
, LW_EXCLUSIVE
);
824 autovac_balance_cost();
825 LWLockRelease(AutovacuumLock
);
827 /* rebuild the list in case the naptime changed */
828 rebuild_database_list(InvalidOid
);
831 /* Process barrier events */
832 if (ProcSignalBarrierPending
)
833 ProcessProcSignalBarrier();
835 /* Perform logging of memory contexts of this process */
836 if (LogMemoryContextPending
)
837 ProcessLogMemoryContextInterrupt();
839 /* Process sinval catchup interrupts that happened while sleeping */
840 ProcessCatchupInterrupt();
844 * Perform a normal exit from the autovac launcher.
847 AutoVacLauncherShutdown(void)
850 (errmsg_internal("autovacuum launcher shutting down")));
851 AutoVacuumShmem
->av_launcherpid
= 0;
853 proc_exit(0); /* done */
857 * Determine the time to sleep, based on the database list.
859 * The "canlaunch" parameter indicates whether we can start a worker right now,
860 * for example due to the workers being all busy. If this is false, we will
861 * cause a long sleep, which will be interrupted when a worker exits.
864 launcher_determine_sleep(bool canlaunch
, bool recursing
, struct timeval
*nap
)
867 * We sleep until the next scheduled vacuum. We trust that when the
868 * database list was built, care was taken so that no entries have times
869 * in the past; if the first entry has too close a next_worker value, or a
870 * time in the past, we will sleep a small nominal time.
874 nap
->tv_sec
= autovacuum_naptime
;
877 else if (!dlist_is_empty(&DatabaseList
))
879 TimestampTz current_time
= GetCurrentTimestamp();
880 TimestampTz next_wakeup
;
885 avdb
= dlist_tail_element(avl_dbase
, adl_node
, &DatabaseList
);
887 next_wakeup
= avdb
->adl_next_worker
;
888 TimestampDifference(current_time
, next_wakeup
, &secs
, &usecs
);
891 nap
->tv_usec
= usecs
;
895 /* list is empty, sleep for whole autovacuum_naptime seconds */
896 nap
->tv_sec
= autovacuum_naptime
;
901 * If the result is exactly zero, it means a database had an entry with
902 * time in the past. Rebuild the list so that the databases are evenly
903 * distributed again, and recalculate the time to sleep. This can happen
904 * if there are more tables needing vacuum than workers, and they all take
905 * longer to vacuum than autovacuum_naptime.
907 * We only recurse once. rebuild_database_list should always return times
908 * in the future, but it seems best not to trust too much on that.
910 if (nap
->tv_sec
== 0 && nap
->tv_usec
== 0 && !recursing
)
912 rebuild_database_list(InvalidOid
);
913 launcher_determine_sleep(canlaunch
, true, nap
);
917 /* The smallest time we'll allow the launcher to sleep. */
918 if (nap
->tv_sec
<= 0 && nap
->tv_usec
<= MIN_AUTOVAC_SLEEPTIME
* 1000)
921 nap
->tv_usec
= MIN_AUTOVAC_SLEEPTIME
* 1000;
925 * If the sleep time is too large, clamp it to an arbitrary maximum (plus
926 * any fractional seconds, for simplicity). This avoids an essentially
927 * infinite sleep in strange cases like the system clock going backwards a
930 if (nap
->tv_sec
> MAX_AUTOVAC_SLEEPTIME
)
931 nap
->tv_sec
= MAX_AUTOVAC_SLEEPTIME
;
935 * Build an updated DatabaseList. It must only contain databases that appear
936 * in pgstats, and must be sorted by next_worker from highest to lowest,
937 * distributed regularly across the next autovacuum_naptime interval.
939 * Receives the Oid of the database that made this list be generated (we call
940 * this the "new" database, because when the database was already present on
941 * the list, we expect that this function is not called at all). The
942 * preexisting list, if any, will be used to preserve the order of the
943 * databases in the autovacuum_naptime period. The new database is put at the
944 * end of the interval. The actual values are not saved, which should not be
948 rebuild_database_list(Oid newdb
)
952 MemoryContext newcxt
;
953 MemoryContext oldcxt
;
954 MemoryContext tmpcxt
;
961 newcxt
= AllocSetContextCreate(AutovacMemCxt
,
962 "Autovacuum database list",
963 ALLOCSET_DEFAULT_SIZES
);
964 tmpcxt
= AllocSetContextCreate(newcxt
,
965 "Autovacuum database list (tmp)",
966 ALLOCSET_DEFAULT_SIZES
);
967 oldcxt
= MemoryContextSwitchTo(tmpcxt
);
970 * Implementing this is not as simple as it sounds, because we need to put
971 * the new database at the end of the list; next the databases that were
972 * already on the list, and finally (at the tail of the list) all the
973 * other databases that are not on the existing list.
975 * To do this, we build an empty hash table of scored databases. We will
976 * start with the lowest score (zero) for the new database, then
977 * increasing scores for the databases in the existing list, in order, and
978 * lastly increasing scores for all databases gotten via
979 * get_database_list() that are not already on the hash.
981 * Then we will put all the hash elements into an array, sort the array by
982 * score, and finally put the array elements into the new doubly linked
985 hctl
.keysize
= sizeof(Oid
);
986 hctl
.entrysize
= sizeof(avl_dbase
);
988 dbhash
= hash_create("autovacuum db hash", 20, &hctl
, /* magic number here
990 HASH_ELEM
| HASH_BLOBS
| HASH_CONTEXT
);
992 /* start by inserting the new database */
994 if (OidIsValid(newdb
))
997 PgStat_StatDBEntry
*entry
;
999 /* only consider this database if it has a pgstat entry */
1000 entry
= pgstat_fetch_stat_dbentry(newdb
);
1003 /* we assume it isn't found because the hash was just created */
1004 db
= hash_search(dbhash
, &newdb
, HASH_ENTER
, NULL
);
1006 /* hash_search already filled in the key */
1007 db
->adl_score
= score
++;
1008 /* next_worker is filled in later */
1012 /* Now insert the databases from the existing list */
1013 dlist_foreach(iter
, &DatabaseList
)
1015 avl_dbase
*avdb
= dlist_container(avl_dbase
, adl_node
, iter
.cur
);
1018 PgStat_StatDBEntry
*entry
;
1021 * skip databases with no stat entries -- in particular, this gets rid
1022 * of dropped databases
1024 entry
= pgstat_fetch_stat_dbentry(avdb
->adl_datid
);
1028 db
= hash_search(dbhash
, &(avdb
->adl_datid
), HASH_ENTER
, &found
);
1032 /* hash_search already filled in the key */
1033 db
->adl_score
= score
++;
1034 /* next_worker is filled in later */
1038 /* finally, insert all qualifying databases not previously inserted */
1039 dblist
= get_database_list();
1040 foreach(cell
, dblist
)
1042 avw_dbase
*avdb
= lfirst(cell
);
1045 PgStat_StatDBEntry
*entry
;
1047 /* only consider databases with a pgstat entry */
1048 entry
= pgstat_fetch_stat_dbentry(avdb
->adw_datid
);
1052 db
= hash_search(dbhash
, &(avdb
->adw_datid
), HASH_ENTER
, &found
);
1053 /* only update the score if the database was not already on the hash */
1056 /* hash_search already filled in the key */
1057 db
->adl_score
= score
++;
1058 /* next_worker is filled in later */
1063 /* from here on, the allocated memory belongs to the new list */
1064 MemoryContextSwitchTo(newcxt
);
1065 dlist_init(&DatabaseList
);
1069 TimestampTz current_time
;
1070 int millis_increment
;
1073 HASH_SEQ_STATUS seq
;
1076 /* put all the hash elements into an array */
1077 dbary
= palloc(nelems
* sizeof(avl_dbase
));
1080 hash_seq_init(&seq
, dbhash
);
1081 while ((db
= hash_seq_search(&seq
)) != NULL
)
1082 memcpy(&(dbary
[i
++]), db
, sizeof(avl_dbase
));
1084 /* sort the array */
1085 qsort(dbary
, nelems
, sizeof(avl_dbase
), db_comparator
);
1088 * Determine the time interval between databases in the schedule. If
1089 * we see that the configured naptime would take us to sleep times
1090 * lower than our min sleep time (which launcher_determine_sleep is
1091 * coded not to allow), silently use a larger naptime (but don't touch
1092 * the GUC variable).
1094 millis_increment
= 1000.0 * autovacuum_naptime
/ nelems
;
1095 if (millis_increment
<= MIN_AUTOVAC_SLEEPTIME
)
1096 millis_increment
= MIN_AUTOVAC_SLEEPTIME
* 1.1;
1098 current_time
= GetCurrentTimestamp();
1101 * move the elements from the array into the dlist, setting the
1102 * next_worker while walking the array
1104 for (i
= 0; i
< nelems
; i
++)
1108 current_time
= TimestampTzPlusMilliseconds(current_time
,
1110 db
->adl_next_worker
= current_time
;
1112 /* later elements should go closer to the head of the list */
1113 dlist_push_head(&DatabaseList
, &db
->adl_node
);
1117 /* all done, clean up memory */
1118 if (DatabaseListCxt
!= NULL
)
1119 MemoryContextDelete(DatabaseListCxt
);
1120 MemoryContextDelete(tmpcxt
);
1121 DatabaseListCxt
= newcxt
;
1122 MemoryContextSwitchTo(oldcxt
);
1125 /* qsort comparator for avl_dbase, using adl_score */
1127 db_comparator(const void *a
, const void *b
)
1129 if (((const avl_dbase
*) a
)->adl_score
== ((const avl_dbase
*) b
)->adl_score
)
1132 return (((const avl_dbase
*) a
)->adl_score
< ((const avl_dbase
*) b
)->adl_score
) ? 1 : -1;
1138 * Bare-bones procedure for starting an autovacuum worker from the launcher.
1139 * It determines what database to work on, sets up shared memory stuff and
1140 * signals postmaster to start the worker. It fails gracefully if invoked when
1141 * autovacuum_workers are already active.
1143 * Return value is the OID of the database that the worker is going to process,
1144 * or InvalidOid if no worker was actually started.
1147 do_start_worker(void)
1151 TransactionId xidForceLimit
;
1152 MultiXactId multiForceLimit
;
1154 bool for_multi_wrap
;
1156 TimestampTz current_time
;
1157 bool skipit
= false;
1158 Oid retval
= InvalidOid
;
1159 MemoryContext tmpcxt
,
1162 /* return quickly when there are no free workers */
1163 LWLockAcquire(AutovacuumLock
, LW_SHARED
);
1164 if (dlist_is_empty(&AutoVacuumShmem
->av_freeWorkers
))
1166 LWLockRelease(AutovacuumLock
);
1169 LWLockRelease(AutovacuumLock
);
1172 * Create and switch to a temporary context to avoid leaking the memory
1173 * allocated for the database list.
1175 tmpcxt
= AllocSetContextCreate(CurrentMemoryContext
,
1176 "Autovacuum start worker (tmp)",
1177 ALLOCSET_DEFAULT_SIZES
);
1178 oldcxt
= MemoryContextSwitchTo(tmpcxt
);
1180 /* Get a list of databases */
1181 dblist
= get_database_list();
1184 * Determine the oldest datfrozenxid/relfrozenxid that we will allow to
1185 * pass without forcing a vacuum. (This limit can be tightened for
1186 * particular tables, but not loosened.)
1188 recentXid
= ReadNextTransactionId();
1189 xidForceLimit
= recentXid
- autovacuum_freeze_max_age
;
1190 /* ensure it's a "normal" XID, else TransactionIdPrecedes misbehaves */
1191 /* this can cause the limit to go backwards by 3, but that's OK */
1192 if (xidForceLimit
< FirstNormalTransactionId
)
1193 xidForceLimit
-= FirstNormalTransactionId
;
1195 /* Also determine the oldest datminmxid we will consider. */
1196 recentMulti
= ReadNextMultiXactId();
1197 multiForceLimit
= recentMulti
- MultiXactMemberFreezeThreshold();
1198 if (multiForceLimit
< FirstMultiXactId
)
1199 multiForceLimit
-= FirstMultiXactId
;
1202 * Choose a database to connect to. We pick the database that was least
1203 * recently auto-vacuumed, or one that needs vacuuming to prevent Xid
1204 * wraparound-related data loss. If any db at risk of Xid wraparound is
1205 * found, we pick the one with oldest datfrozenxid, independently of
1206 * autovacuum times; similarly we pick the one with the oldest datminmxid
1207 * if any is in MultiXactId wraparound. Note that those in Xid wraparound
1208 * danger are given more priority than those in multi wraparound danger.
1210 * Note that a database with no stats entry is not considered, except for
1211 * Xid wraparound purposes. The theory is that if no one has ever
1212 * connected to it since the stats were last initialized, it doesn't need
1215 * XXX This could be improved if we had more info about whether it needs
1216 * vacuuming before connecting to it. Perhaps look through the pgstats
1217 * data for the database's tables? One idea is to keep track of the
1218 * number of new and dead tuples per database in pgstats. However it
1219 * isn't clear how to construct a metric that measures that and not cause
1220 * starvation for less busy databases.
1223 for_xid_wrap
= false;
1224 for_multi_wrap
= false;
1225 current_time
= GetCurrentTimestamp();
1226 foreach(cell
, dblist
)
1228 avw_dbase
*tmp
= lfirst(cell
);
1231 /* Check to see if this one is at risk of wraparound */
1232 if (TransactionIdPrecedes(tmp
->adw_frozenxid
, xidForceLimit
))
1235 TransactionIdPrecedes(tmp
->adw_frozenxid
,
1236 avdb
->adw_frozenxid
))
1238 for_xid_wrap
= true;
1241 else if (for_xid_wrap
)
1242 continue; /* ignore not-at-risk DBs */
1243 else if (MultiXactIdPrecedes(tmp
->adw_minmulti
, multiForceLimit
))
1246 MultiXactIdPrecedes(tmp
->adw_minmulti
, avdb
->adw_minmulti
))
1248 for_multi_wrap
= true;
1251 else if (for_multi_wrap
)
1252 continue; /* ignore not-at-risk DBs */
1254 /* Find pgstat entry if any */
1255 tmp
->adw_entry
= pgstat_fetch_stat_dbentry(tmp
->adw_datid
);
1258 * Skip a database with no pgstat entry; it means it hasn't seen any
1261 if (!tmp
->adw_entry
)
1265 * Also, skip a database that appears on the database list as having
1266 * been processed recently (less than autovacuum_naptime seconds ago).
1267 * We do this so that we don't select a database which we just
1268 * selected, but that pgstat hasn't gotten around to updating the last
1269 * autovacuum time yet.
1273 dlist_reverse_foreach(iter
, &DatabaseList
)
1275 avl_dbase
*dbp
= dlist_container(avl_dbase
, adl_node
, iter
.cur
);
1277 if (dbp
->adl_datid
== tmp
->adw_datid
)
1280 * Skip this database if its next_worker value falls between
1281 * the current time and the current time plus naptime.
1283 if (!TimestampDifferenceExceeds(dbp
->adl_next_worker
,
1285 !TimestampDifferenceExceeds(current_time
,
1286 dbp
->adl_next_worker
,
1287 autovacuum_naptime
* 1000))
1297 * Remember the db with oldest autovac time. (If we are here, both
1298 * tmp->entry and db->entry must be non-null.)
1301 tmp
->adw_entry
->last_autovac_time
< avdb
->adw_entry
->last_autovac_time
)
1305 /* Found a database -- process it */
1311 LWLockAcquire(AutovacuumLock
, LW_EXCLUSIVE
);
1314 * Get a worker entry from the freelist. We checked above, so there
1315 * really should be a free slot.
1317 wptr
= dlist_pop_head_node(&AutoVacuumShmem
->av_freeWorkers
);
1319 worker
= dlist_container(WorkerInfoData
, wi_links
, wptr
);
1320 worker
->wi_dboid
= avdb
->adw_datid
;
1321 worker
->wi_proc
= NULL
;
1322 worker
->wi_launchtime
= GetCurrentTimestamp();
1324 AutoVacuumShmem
->av_startingWorker
= worker
;
1326 LWLockRelease(AutovacuumLock
);
1328 SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_WORKER
);
1330 retval
= avdb
->adw_datid
;
1335 * If we skipped all databases on the list, rebuild it, because it
1336 * probably contains a dropped database.
1338 rebuild_database_list(InvalidOid
);
1341 MemoryContextSwitchTo(oldcxt
);
1342 MemoryContextDelete(tmpcxt
);
1350 * Wrapper for starting a worker from the launcher. Besides actually starting
1351 * it, update the database list to reflect the next time that another one will
1352 * need to be started on the selected database. The actual database choice is
1353 * left to do_start_worker.
1355 * This routine is also expected to insert an entry into the database list if
1356 * the selected database was previously absent from the list.
1359 launch_worker(TimestampTz now
)
1364 dbid
= do_start_worker();
1365 if (OidIsValid(dbid
))
1370 * Walk the database list and update the corresponding entry. If the
1371 * database is not on the list, we'll recreate the list.
1373 dlist_foreach(iter
, &DatabaseList
)
1375 avl_dbase
*avdb
= dlist_container(avl_dbase
, adl_node
, iter
.cur
);
1377 if (avdb
->adl_datid
== dbid
)
1382 * add autovacuum_naptime seconds to the current time, and use
1383 * that as the new "next_worker" field for this database.
1385 avdb
->adl_next_worker
=
1386 TimestampTzPlusMilliseconds(now
, autovacuum_naptime
* 1000);
1388 dlist_move_head(&DatabaseList
, iter
.cur
);
1394 * If the database was not present in the database list, we rebuild
1395 * the list. It's possible that the database does not get into the
1396 * list anyway, for example if it's a database that doesn't have a
1397 * pgstat entry, but this is not a problem because we don't want to
1398 * schedule workers regularly into those in any case.
1401 rebuild_database_list(dbid
);
1406 * Called from postmaster to signal a failure to fork a process to become
1407 * worker. The postmaster should kill(SIGUSR2) the launcher shortly
1408 * after calling this function.
1411 AutoVacWorkerFailed(void)
1413 AutoVacuumShmem
->av_signal
[AutoVacForkFailed
] = true;
1416 /* SIGUSR2: a worker is up and running, or just finished, or failed to fork */
1418 avl_sigusr2_handler(SIGNAL_ARGS
)
1420 int save_errno
= errno
;
1429 /********************************************************************
1430 * AUTOVACUUM WORKER CODE
1431 ********************************************************************/
1435 * forkexec routines for the autovacuum worker.
1437 * Format up the arglist, then fork and exec.
1440 avworker_forkexec(void)
1445 av
[ac
++] = "postgres";
1446 av
[ac
++] = "--forkavworker";
1447 av
[ac
++] = NULL
; /* filled in by postmaster_forkexec */
1450 Assert(ac
< lengthof(av
));
1452 return postmaster_forkexec(ac
, av
);
1456 * We need this set from the outside, before InitProcess is called
1459 AutovacuumWorkerIAm(void)
1461 am_autovacuum_worker
= true;
1466 * Main entry point for autovacuum worker process.
1468 * This code is heavily based on pgarch.c, q.v.
1471 StartAutoVacWorker(void)
1476 switch ((worker_pid
= avworker_forkexec()))
1478 switch ((worker_pid
= fork_process()))
1483 (errmsg("could not fork autovacuum worker process: %m")));
1486 #ifndef EXEC_BACKEND
1488 /* in postmaster child ... */
1489 InitPostmasterChild();
1491 /* Close the postmaster's sockets */
1492 ClosePostmasterPorts(false);
1494 AutoVacWorkerMain(0, NULL
);
1498 return (int) worker_pid
;
1501 /* shouldn't get here */
1508 NON_EXEC_STATIC
void
1509 AutoVacWorkerMain(int argc
, char *argv
[])
1511 sigjmp_buf local_sigjmp_buf
;
1514 am_autovacuum_worker
= true;
1516 MyBackendType
= B_AUTOVAC_WORKER
;
1517 init_ps_display(NULL
);
1519 SetProcessingMode(InitProcessing
);
1522 * Set up signal handlers. We operate on databases much like a regular
1523 * backend, so we use the same signal handling. See equivalent code in
1526 pqsignal(SIGHUP
, SignalHandlerForConfigReload
);
1529 * SIGINT is used to signal canceling the current table's vacuum; SIGTERM
1530 * means abort and exit cleanly, and SIGQUIT means abandon ship.
1532 pqsignal(SIGINT
, StatementCancelHandler
);
1533 pqsignal(SIGTERM
, die
);
1534 /* SIGQUIT handler was already set up by InitPostmasterChild */
1536 InitializeTimeouts(); /* establishes SIGALRM handler */
1538 pqsignal(SIGPIPE
, SIG_IGN
);
1539 pqsignal(SIGUSR1
, procsignal_sigusr1_handler
);
1540 pqsignal(SIGUSR2
, SIG_IGN
);
1541 pqsignal(SIGFPE
, FloatExceptionHandler
);
1542 pqsignal(SIGCHLD
, SIG_DFL
);
1545 * Create a per-backend PGPROC struct in shared memory, except in the
1546 * EXEC_BACKEND case where this was done in SubPostmasterMain. We must do
1547 * this before we can use LWLocks (and in the EXEC_BACKEND case we already
1548 * had to do some stuff with LWLocks).
1550 #ifndef EXEC_BACKEND
1554 /* Early initialization */
1558 * If an exception is encountered, processing resumes here.
1560 * Unlike most auxiliary processes, we don't attempt to continue
1561 * processing after an error; we just clean up and exit. The autovac
1562 * launcher is responsible for spawning another worker later.
1564 * Note that we use sigsetjmp(..., 1), so that the prevailing signal mask
1565 * (to wit, BlockSig) will be restored when longjmp'ing to here. Thus,
1566 * signals other than SIGQUIT will be blocked until we exit. It might
1567 * seem that this policy makes the HOLD_INTERRUPTS() call redundant, but
1568 * it is not since InterruptPending might be set already.
1570 if (sigsetjmp(local_sigjmp_buf
, 1) != 0)
1572 /* since not using PG_TRY, must reset error stack by hand */
1573 error_context_stack
= NULL
;
1575 /* Prevents interrupts while cleaning up */
1578 /* Report the error to the server log */
1582 * We can now go away. Note that because we called InitProcess, a
1583 * callback was registered to do ProcKill, which will clean up
1589 /* We can now handle ereport(ERROR) */
1590 PG_exception_stack
= &local_sigjmp_buf
;
1592 sigprocmask(SIG_SETMASK
, &UnBlockSig
, NULL
);
1595 * Set always-secure search path, so malicious users can't redirect user
1596 * code (e.g. pg_index.indexprs). (That code runs in a
1597 * SECURITY_RESTRICTED_OPERATION sandbox, so malicious users could not
1598 * take control of the entire autovacuum worker in any case.)
1600 SetConfigOption("search_path", "", PGC_SUSET
, PGC_S_OVERRIDE
);
1603 * Force zero_damaged_pages OFF in the autovac process, even if it is set
1604 * in postgresql.conf. We don't really want such a dangerous option being
1605 * applied non-interactively.
1607 SetConfigOption("zero_damaged_pages", "false", PGC_SUSET
, PGC_S_OVERRIDE
);
1610 * Force settable timeouts off to avoid letting these settings prevent
1611 * regular maintenance from being executed.
1613 SetConfigOption("statement_timeout", "0", PGC_SUSET
, PGC_S_OVERRIDE
);
1614 SetConfigOption("lock_timeout", "0", PGC_SUSET
, PGC_S_OVERRIDE
);
1615 SetConfigOption("idle_in_transaction_session_timeout", "0",
1616 PGC_SUSET
, PGC_S_OVERRIDE
);
1619 * Force default_transaction_isolation to READ COMMITTED. We don't want
1620 * to pay the overhead of serializable mode, nor add any risk of causing
1621 * deadlocks or delaying other transactions.
1623 SetConfigOption("default_transaction_isolation", "read committed",
1624 PGC_SUSET
, PGC_S_OVERRIDE
);
1627 * Force synchronous replication off to allow regular maintenance even if
1628 * we are waiting for standbys to connect. This is important to ensure we
1629 * aren't blocked from performing anti-wraparound tasks.
1631 if (synchronous_commit
> SYNCHRONOUS_COMMIT_LOCAL_FLUSH
)
1632 SetConfigOption("synchronous_commit", "local",
1633 PGC_SUSET
, PGC_S_OVERRIDE
);
1636 * Even when system is configured to use a different fetch consistency,
1637 * for autovac we always want fresh stats.
1639 SetConfigOption("stats_fetch_consistency", "none", PGC_SUSET
, PGC_S_OVERRIDE
);
1642 * Get the info about the database we're going to work on.
1644 LWLockAcquire(AutovacuumLock
, LW_EXCLUSIVE
);
1647 * beware of startingWorker being INVALID; this should normally not
1648 * happen, but if a worker fails after forking and before this, the
1649 * launcher might have decided to remove it from the queue and start
1652 if (AutoVacuumShmem
->av_startingWorker
!= NULL
)
1654 MyWorkerInfo
= AutoVacuumShmem
->av_startingWorker
;
1655 dbid
= MyWorkerInfo
->wi_dboid
;
1656 MyWorkerInfo
->wi_proc
= MyProc
;
1658 /* insert into the running list */
1659 dlist_push_head(&AutoVacuumShmem
->av_runningWorkers
,
1660 &MyWorkerInfo
->wi_links
);
1663 * remove from the "starting" pointer, so that the launcher can start
1664 * a new worker if required
1666 AutoVacuumShmem
->av_startingWorker
= NULL
;
1667 LWLockRelease(AutovacuumLock
);
1669 on_shmem_exit(FreeWorkerInfo
, 0);
1671 /* wake up the launcher */
1672 if (AutoVacuumShmem
->av_launcherpid
!= 0)
1673 kill(AutoVacuumShmem
->av_launcherpid
, SIGUSR2
);
1677 /* no worker entry for me, go away */
1678 elog(WARNING
, "autovacuum worker started without a worker entry");
1680 LWLockRelease(AutovacuumLock
);
1683 if (OidIsValid(dbid
))
1685 char dbname
[NAMEDATALEN
];
1688 * Report autovac startup to the cumulative stats system. We
1689 * deliberately do this before InitPostgres, so that the
1690 * last_autovac_time will get updated even if the connection attempt
1691 * fails. This is to prevent autovac from getting "stuck" repeatedly
1692 * selecting an unopenable database, rather than making any progress
1693 * on stuff it can connect to.
1695 pgstat_report_autovac(dbid
);
1698 * Connect to the selected database, specifying no particular user
1700 * Note: if we have selected a just-deleted database (due to using
1701 * stale stats info), we'll fail and exit here.
1703 InitPostgres(NULL
, dbid
, NULL
, InvalidOid
, false, false,
1705 SetProcessingMode(NormalProcessing
);
1706 set_ps_display(dbname
);
1708 (errmsg_internal("autovacuum: processing database \"%s\"", dbname
)));
1711 pg_usleep(PostAuthDelay
* 1000000L);
1713 /* And do an appropriate amount of work */
1714 recentXid
= ReadNextTransactionId();
1715 recentMulti
= ReadNextMultiXactId();
1720 * The launcher will be notified of my death in ProcKill, *if* we managed
1721 * to get a worker slot at all
1724 /* All done, go away */
1729 * Return a WorkerInfo to the free list
1732 FreeWorkerInfo(int code
, Datum arg
)
1734 if (MyWorkerInfo
!= NULL
)
1736 LWLockAcquire(AutovacuumLock
, LW_EXCLUSIVE
);
1739 * Wake the launcher up so that he can launch a new worker immediately
1740 * if required. We only save the launcher's PID in local memory here;
1741 * the actual signal will be sent when the PGPROC is recycled. Note
1742 * that we always do this, so that the launcher can rebalance the cost
1743 * limit setting of the remaining workers.
1745 * We somewhat ignore the risk that the launcher changes its PID
1746 * between us reading it and the actual kill; we expect ProcKill to be
1747 * called shortly after us, and we assume that PIDs are not reused too
1748 * quickly after a process exits.
1750 AutovacuumLauncherPid
= AutoVacuumShmem
->av_launcherpid
;
1752 dlist_delete(&MyWorkerInfo
->wi_links
);
1753 MyWorkerInfo
->wi_dboid
= InvalidOid
;
1754 MyWorkerInfo
->wi_tableoid
= InvalidOid
;
1755 MyWorkerInfo
->wi_sharedrel
= false;
1756 MyWorkerInfo
->wi_proc
= NULL
;
1757 MyWorkerInfo
->wi_launchtime
= 0;
1758 MyWorkerInfo
->wi_dobalance
= false;
1759 MyWorkerInfo
->wi_cost_delay
= 0;
1760 MyWorkerInfo
->wi_cost_limit
= 0;
1761 MyWorkerInfo
->wi_cost_limit_base
= 0;
1762 dlist_push_head(&AutoVacuumShmem
->av_freeWorkers
,
1763 &MyWorkerInfo
->wi_links
);
1764 /* not mine anymore */
1765 MyWorkerInfo
= NULL
;
1768 * now that we're inactive, cause a rebalancing of the surviving
1771 AutoVacuumShmem
->av_signal
[AutoVacRebalance
] = true;
1772 LWLockRelease(AutovacuumLock
);
1777 * Update the cost-based delay parameters, so that multiple workers consume
1778 * each a fraction of the total available I/O.
1781 AutoVacuumUpdateDelay(void)
1785 VacuumCostDelay
= MyWorkerInfo
->wi_cost_delay
;
1786 VacuumCostLimit
= MyWorkerInfo
->wi_cost_limit
;
1791 * autovac_balance_cost
1792 * Recalculate the cost limit setting for each active worker.
1794 * Caller must hold the AutovacuumLock in exclusive mode.
1797 autovac_balance_cost(void)
1800 * The idea here is that we ration out I/O equally. The amount of I/O
1801 * that a worker can consume is determined by cost_limit/cost_delay, so we
1802 * try to equalize those ratios rather than the raw limit settings.
1804 * note: in cost_limit, zero also means use value from elsewhere, because
1805 * zero is not a valid value.
1807 int vac_cost_limit
= (autovacuum_vac_cost_limit
> 0 ?
1808 autovacuum_vac_cost_limit
: VacuumCostLimit
);
1809 double vac_cost_delay
= (autovacuum_vac_cost_delay
>= 0 ?
1810 autovacuum_vac_cost_delay
: VacuumCostDelay
);
1815 /* not set? nothing to do */
1816 if (vac_cost_limit
<= 0 || vac_cost_delay
<= 0)
1819 /* calculate the total base cost limit of participating active workers */
1821 dlist_foreach(iter
, &AutoVacuumShmem
->av_runningWorkers
)
1823 WorkerInfo worker
= dlist_container(WorkerInfoData
, wi_links
, iter
.cur
);
1825 if (worker
->wi_proc
!= NULL
&&
1826 worker
->wi_dobalance
&&
1827 worker
->wi_cost_limit_base
> 0 && worker
->wi_cost_delay
> 0)
1829 (double) worker
->wi_cost_limit_base
/ worker
->wi_cost_delay
;
1832 /* there are no cost limits -- nothing to do */
1833 if (cost_total
<= 0)
1837 * Adjust cost limit of each active worker to balance the total of cost
1838 * limit to autovacuum_vacuum_cost_limit.
1840 cost_avail
= (double) vac_cost_limit
/ vac_cost_delay
;
1841 dlist_foreach(iter
, &AutoVacuumShmem
->av_runningWorkers
)
1843 WorkerInfo worker
= dlist_container(WorkerInfoData
, wi_links
, iter
.cur
);
1845 if (worker
->wi_proc
!= NULL
&&
1846 worker
->wi_dobalance
&&
1847 worker
->wi_cost_limit_base
> 0 && worker
->wi_cost_delay
> 0)
1850 (cost_avail
* worker
->wi_cost_limit_base
/ cost_total
);
1853 * We put a lower bound of 1 on the cost_limit, to avoid division-
1854 * by-zero in the vacuum code. Also, in case of roundoff trouble
1855 * in these calculations, let's be sure we don't ever set
1856 * cost_limit to more than the base value.
1858 worker
->wi_cost_limit
= Max(Min(limit
,
1859 worker
->wi_cost_limit_base
),
1863 if (worker
->wi_proc
!= NULL
)
1864 elog(DEBUG2
, "autovac_balance_cost(pid=%d db=%u, rel=%u, dobalance=%s cost_limit=%d, cost_limit_base=%d, cost_delay=%g)",
1865 worker
->wi_proc
->pid
, worker
->wi_dboid
, worker
->wi_tableoid
,
1866 worker
->wi_dobalance
? "yes" : "no",
1867 worker
->wi_cost_limit
, worker
->wi_cost_limit_base
,
1868 worker
->wi_cost_delay
);
1874 * Return a list of all databases found in pg_database.
1876 * The list and associated data is allocated in the caller's memory context,
1877 * which is in charge of ensuring that it's properly cleaned up afterwards.
1879 * Note: this is the only function in which the autovacuum launcher uses a
1880 * transaction. Although we aren't attached to any particular database and
1881 * therefore can't access most catalogs, we do have enough infrastructure
1882 * to do a seqscan on pg_database.
1885 get_database_list(void)
1891 MemoryContext resultcxt
;
1893 /* This is the context that we will allocate our output data in */
1894 resultcxt
= CurrentMemoryContext
;
1897 * Start a transaction so we can access pg_database, and get a snapshot.
1898 * We don't have a use for the snapshot itself, but we're interested in
1899 * the secondary effect that it sets RecentGlobalXmin. (This is critical
1900 * for anything that reads heap pages, because HOT may decide to prune
1901 * them even if the process doesn't attempt to modify any tuples.)
1903 * FIXME: This comment is inaccurate / the code buggy. A snapshot that is
1904 * not pushed/active does not reliably prevent HOT pruning (->xmin could
1905 * e.g. be cleared when cache invalidations are processed).
1907 StartTransactionCommand();
1908 (void) GetTransactionSnapshot();
1910 rel
= table_open(DatabaseRelationId
, AccessShareLock
);
1911 scan
= table_beginscan_catalog(rel
, 0, NULL
);
1913 while (HeapTupleIsValid(tup
= heap_getnext(scan
, ForwardScanDirection
)))
1915 Form_pg_database pgdatabase
= (Form_pg_database
) GETSTRUCT(tup
);
1917 MemoryContext oldcxt
;
1920 * Allocate our results in the caller's context, not the
1921 * transaction's. We do this inside the loop, and restore the original
1922 * context at the end, so that leaky things like heap_getnext() are
1923 * not called in a potentially long-lived context.
1925 oldcxt
= MemoryContextSwitchTo(resultcxt
);
1927 avdb
= (avw_dbase
*) palloc(sizeof(avw_dbase
));
1929 avdb
->adw_datid
= pgdatabase
->oid
;
1930 avdb
->adw_name
= pstrdup(NameStr(pgdatabase
->datname
));
1931 avdb
->adw_frozenxid
= pgdatabase
->datfrozenxid
;
1932 avdb
->adw_minmulti
= pgdatabase
->datminmxid
;
1933 /* this gets set later: */
1934 avdb
->adw_entry
= NULL
;
1936 dblist
= lappend(dblist
, avdb
);
1937 MemoryContextSwitchTo(oldcxt
);
1940 table_endscan(scan
);
1941 table_close(rel
, AccessShareLock
);
1943 CommitTransactionCommand();
1945 /* Be sure to restore caller's memory context */
1946 MemoryContextSwitchTo(resultcxt
);
1952 * Process a database table-by-table
1954 * Note that CHECK_FOR_INTERRUPTS is supposed to be used in certain spots in
1955 * order not to ignore shutdown commands for too long.
1962 TableScanDesc relScan
;
1963 Form_pg_database dbForm
;
1964 List
*table_oids
= NIL
;
1965 List
*orphan_oids
= NIL
;
1967 HTAB
*table_toast_map
;
1968 ListCell
*volatile cell
;
1969 BufferAccessStrategy bstrategy
;
1971 TupleDesc pg_class_desc
;
1972 int effective_multixact_freeze_max_age
;
1973 bool did_vacuum
= false;
1974 bool found_concurrent_worker
= false;
1978 * StartTransactionCommand and CommitTransactionCommand will automatically
1979 * switch to other contexts. We need this one to keep the list of
1980 * relations to vacuum/analyze across transactions.
1982 AutovacMemCxt
= AllocSetContextCreate(TopMemoryContext
,
1983 "Autovacuum worker",
1984 ALLOCSET_DEFAULT_SIZES
);
1985 MemoryContextSwitchTo(AutovacMemCxt
);
1987 /* Start a transaction so our commands have one to play into. */
1988 StartTransactionCommand();
1991 * Compute the multixact age for which freezing is urgent. This is
1992 * normally autovacuum_multixact_freeze_max_age, but may be less if we are
1993 * short of multixact member space.
1995 effective_multixact_freeze_max_age
= MultiXactMemberFreezeThreshold();
1998 * Find the pg_database entry and select the default freeze ages. We use
1999 * zero in template and nonconnectable databases, else the system-wide
2002 tuple
= SearchSysCache1(DATABASEOID
, ObjectIdGetDatum(MyDatabaseId
));
2003 if (!HeapTupleIsValid(tuple
))
2004 elog(ERROR
, "cache lookup failed for database %u", MyDatabaseId
);
2005 dbForm
= (Form_pg_database
) GETSTRUCT(tuple
);
2007 if (dbForm
->datistemplate
|| !dbForm
->datallowconn
)
2009 default_freeze_min_age
= 0;
2010 default_freeze_table_age
= 0;
2011 default_multixact_freeze_min_age
= 0;
2012 default_multixact_freeze_table_age
= 0;
2016 default_freeze_min_age
= vacuum_freeze_min_age
;
2017 default_freeze_table_age
= vacuum_freeze_table_age
;
2018 default_multixact_freeze_min_age
= vacuum_multixact_freeze_min_age
;
2019 default_multixact_freeze_table_age
= vacuum_multixact_freeze_table_age
;
2022 ReleaseSysCache(tuple
);
2024 /* StartTransactionCommand changed elsewhere */
2025 MemoryContextSwitchTo(AutovacMemCxt
);
2027 classRel
= table_open(RelationRelationId
, AccessShareLock
);
2029 /* create a copy so we can use it after closing pg_class */
2030 pg_class_desc
= CreateTupleDescCopy(RelationGetDescr(classRel
));
2032 /* create hash table for toast <-> main relid mapping */
2033 ctl
.keysize
= sizeof(Oid
);
2034 ctl
.entrysize
= sizeof(av_relation
);
2036 table_toast_map
= hash_create("TOAST to main relid map",
2039 HASH_ELEM
| HASH_BLOBS
);
2042 * Scan pg_class to determine which tables to vacuum.
2044 * We do this in two passes: on the first one we collect the list of plain
2045 * relations and materialized views, and on the second one we collect
2046 * TOAST tables. The reason for doing the second pass is that during it we
2047 * want to use the main relation's pg_class.reloptions entry if the TOAST
2048 * table does not have any, and we cannot obtain it unless we know
2049 * beforehand what's the main table OID.
2051 * We need to check TOAST tables separately because in cases with short,
2052 * wide tables there might be proportionally much more activity in the
2053 * TOAST table than in its parent.
2055 relScan
= table_beginscan_catalog(classRel
, 0, NULL
);
2058 * On the first pass, we collect main tables to vacuum, and also the main
2059 * table relid to TOAST relid mapping.
2061 while ((tuple
= heap_getnext(relScan
, ForwardScanDirection
)) != NULL
)
2063 Form_pg_class classForm
= (Form_pg_class
) GETSTRUCT(tuple
);
2064 PgStat_StatTabEntry
*tabentry
;
2065 AutoVacOpts
*relopts
;
2071 if (classForm
->relkind
!= RELKIND_RELATION
&&
2072 classForm
->relkind
!= RELKIND_MATVIEW
)
2075 relid
= classForm
->oid
;
2078 * Check if it is a temp table (presumably, of some other backend's).
2079 * We cannot safely process other backends' temp tables.
2081 if (classForm
->relpersistence
== RELPERSISTENCE_TEMP
)
2084 * We just ignore it if the owning backend is still active and
2085 * using the temporary schema. Also, for safety, ignore it if the
2086 * namespace doesn't exist or isn't a temp namespace after all.
2088 if (checkTempNamespaceStatus(classForm
->relnamespace
) == TEMP_NAMESPACE_IDLE
)
2091 * The table seems to be orphaned -- although it might be that
2092 * the owning backend has already deleted it and exited; our
2093 * pg_class scan snapshot is not necessarily up-to-date
2094 * anymore, so we could be looking at a committed-dead entry.
2095 * Remember it so we can try to delete it later.
2097 orphan_oids
= lappend_oid(orphan_oids
, relid
);
2102 /* Fetch reloptions and the pgstat entry for this table */
2103 relopts
= extract_autovac_opts(tuple
, pg_class_desc
);
2104 tabentry
= pgstat_fetch_stat_tabentry_ext(classForm
->relisshared
,
2107 /* Check if it needs vacuum or analyze */
2108 relation_needs_vacanalyze(relid
, relopts
, classForm
, tabentry
,
2109 effective_multixact_freeze_max_age
,
2110 &dovacuum
, &doanalyze
, &wraparound
);
2112 /* Relations that need work are added to table_oids */
2113 if (dovacuum
|| doanalyze
)
2114 table_oids
= lappend_oid(table_oids
, relid
);
2117 * Remember TOAST associations for the second pass. Note: we must do
2118 * this whether or not the table is going to be vacuumed, because we
2119 * don't automatically vacuum toast tables along the parent table.
2121 if (OidIsValid(classForm
->reltoastrelid
))
2123 av_relation
*hentry
;
2126 hentry
= hash_search(table_toast_map
,
2127 &classForm
->reltoastrelid
,
2128 HASH_ENTER
, &found
);
2132 /* hash_search already filled in the key */
2133 hentry
->ar_relid
= relid
;
2134 hentry
->ar_hasrelopts
= false;
2135 if (relopts
!= NULL
)
2137 hentry
->ar_hasrelopts
= true;
2138 memcpy(&hentry
->ar_reloptions
, relopts
,
2139 sizeof(AutoVacOpts
));
2145 table_endscan(relScan
);
2147 /* second pass: check TOAST tables */
2149 Anum_pg_class_relkind
,
2150 BTEqualStrategyNumber
, F_CHAREQ
,
2151 CharGetDatum(RELKIND_TOASTVALUE
));
2153 relScan
= table_beginscan_catalog(classRel
, 1, &key
);
2154 while ((tuple
= heap_getnext(relScan
, ForwardScanDirection
)) != NULL
)
2156 Form_pg_class classForm
= (Form_pg_class
) GETSTRUCT(tuple
);
2157 PgStat_StatTabEntry
*tabentry
;
2159 AutoVacOpts
*relopts
= NULL
;
2165 * We cannot safely process other backends' temp tables, so skip 'em.
2167 if (classForm
->relpersistence
== RELPERSISTENCE_TEMP
)
2170 relid
= classForm
->oid
;
2173 * fetch reloptions -- if this toast table does not have them, try the
2176 relopts
= extract_autovac_opts(tuple
, pg_class_desc
);
2177 if (relopts
== NULL
)
2179 av_relation
*hentry
;
2182 hentry
= hash_search(table_toast_map
, &relid
, HASH_FIND
, &found
);
2183 if (found
&& hentry
->ar_hasrelopts
)
2184 relopts
= &hentry
->ar_reloptions
;
2187 /* Fetch the pgstat entry for this table */
2188 tabentry
= pgstat_fetch_stat_tabentry_ext(classForm
->relisshared
,
2191 relation_needs_vacanalyze(relid
, relopts
, classForm
, tabentry
,
2192 effective_multixact_freeze_max_age
,
2193 &dovacuum
, &doanalyze
, &wraparound
);
2195 /* ignore analyze for toast tables */
2197 table_oids
= lappend_oid(table_oids
, relid
);
2200 table_endscan(relScan
);
2201 table_close(classRel
, AccessShareLock
);
2204 * Recheck orphan temporary tables, and if they still seem orphaned, drop
2205 * them. We'll eat a transaction per dropped table, which might seem
2206 * excessive, but we should only need to do anything as a result of a
2207 * previous backend crash, so this should not happen often enough to
2208 * justify "optimizing". Using separate transactions ensures that we
2209 * don't bloat the lock table if there are many temp tables to be dropped,
2210 * and it ensures that we don't lose work if a deletion attempt fails.
2212 foreach(cell
, orphan_oids
)
2214 Oid relid
= lfirst_oid(cell
);
2215 Form_pg_class classForm
;
2216 ObjectAddress object
;
2219 * Check for user-requested abort.
2221 CHECK_FOR_INTERRUPTS();
2224 * Try to lock the table. If we can't get the lock immediately,
2225 * somebody else is using (or dropping) the table, so it's not our
2226 * concern anymore. Having the lock prevents race conditions below.
2228 if (!ConditionalLockRelationOid(relid
, AccessExclusiveLock
))
2232 * Re-fetch the pg_class tuple and re-check whether it still seems to
2233 * be an orphaned temp table. If it's not there or no longer the same
2234 * relation, ignore it.
2236 tuple
= SearchSysCacheCopy1(RELOID
, ObjectIdGetDatum(relid
));
2237 if (!HeapTupleIsValid(tuple
))
2239 /* be sure to drop useless lock so we don't bloat lock table */
2240 UnlockRelationOid(relid
, AccessExclusiveLock
);
2243 classForm
= (Form_pg_class
) GETSTRUCT(tuple
);
2246 * Make all the same tests made in the loop above. In event of OID
2247 * counter wraparound, the pg_class entry we have now might be
2248 * completely unrelated to the one we saw before.
2250 if (!((classForm
->relkind
== RELKIND_RELATION
||
2251 classForm
->relkind
== RELKIND_MATVIEW
) &&
2252 classForm
->relpersistence
== RELPERSISTENCE_TEMP
))
2254 UnlockRelationOid(relid
, AccessExclusiveLock
);
2258 if (checkTempNamespaceStatus(classForm
->relnamespace
) != TEMP_NAMESPACE_IDLE
)
2260 UnlockRelationOid(relid
, AccessExclusiveLock
);
2264 /* OK, let's delete it */
2266 (errmsg("autovacuum: dropping orphan temp table \"%s.%s.%s\"",
2267 get_database_name(MyDatabaseId
),
2268 get_namespace_name(classForm
->relnamespace
),
2269 NameStr(classForm
->relname
))));
2271 object
.classId
= RelationRelationId
;
2272 object
.objectId
= relid
;
2273 object
.objectSubId
= 0;
2274 performDeletion(&object
, DROP_CASCADE
,
2275 PERFORM_DELETION_INTERNAL
|
2276 PERFORM_DELETION_QUIETLY
|
2277 PERFORM_DELETION_SKIP_EXTENSIONS
);
2280 * To commit the deletion, end current transaction and start a new
2281 * one. Note this also releases the lock we took.
2283 CommitTransactionCommand();
2284 StartTransactionCommand();
2286 /* StartTransactionCommand changed current memory context */
2287 MemoryContextSwitchTo(AutovacMemCxt
);
2291 * Create a buffer access strategy object for VACUUM to use. We want to
2292 * use the same one across all the vacuum operations we perform, since the
2293 * point is for VACUUM not to blow out the shared cache.
2295 bstrategy
= GetAccessStrategy(BAS_VACUUM
);
2298 * create a memory context to act as fake PortalContext, so that the
2299 * contexts created in the vacuum code are cleaned up for each table.
2301 PortalContext
= AllocSetContextCreate(AutovacMemCxt
,
2302 "Autovacuum Portal",
2303 ALLOCSET_DEFAULT_SIZES
);
2306 * Perform operations on collected tables.
2308 foreach(cell
, table_oids
)
2310 Oid relid
= lfirst_oid(cell
);
2315 double stdVacuumCostDelay
;
2316 int stdVacuumCostLimit
;
2319 CHECK_FOR_INTERRUPTS();
2322 * Check for config changes before processing each collected table.
2324 if (ConfigReloadPending
)
2326 ConfigReloadPending
= false;
2327 ProcessConfigFile(PGC_SIGHUP
);
2330 * You might be tempted to bail out if we see autovacuum is now
2331 * disabled. Must resist that temptation -- this might be a
2332 * for-wraparound emergency worker, in which case that would be
2333 * entirely inappropriate.
2338 * Find out whether the table is shared or not. (It's slightly
2339 * annoying to fetch the syscache entry just for this, but in typical
2340 * cases it adds little cost because table_recheck_autovac would
2341 * refetch the entry anyway. We could buy that back by copying the
2342 * tuple here and passing it to table_recheck_autovac, but that
2343 * increases the odds of that function working with stale data.)
2345 classTup
= SearchSysCache1(RELOID
, ObjectIdGetDatum(relid
));
2346 if (!HeapTupleIsValid(classTup
))
2347 continue; /* somebody deleted the rel, forget it */
2348 isshared
= ((Form_pg_class
) GETSTRUCT(classTup
))->relisshared
;
2349 ReleaseSysCache(classTup
);
2352 * Hold schedule lock from here until we've claimed the table. We
2353 * also need the AutovacuumLock to walk the worker array, but that one
2354 * can just be a shared lock.
2356 LWLockAcquire(AutovacuumScheduleLock
, LW_EXCLUSIVE
);
2357 LWLockAcquire(AutovacuumLock
, LW_SHARED
);
2360 * Check whether the table is being vacuumed concurrently by another
2364 dlist_foreach(iter
, &AutoVacuumShmem
->av_runningWorkers
)
2366 WorkerInfo worker
= dlist_container(WorkerInfoData
, wi_links
, iter
.cur
);
2369 if (worker
== MyWorkerInfo
)
2372 /* ignore workers in other databases (unless table is shared) */
2373 if (!worker
->wi_sharedrel
&& worker
->wi_dboid
!= MyDatabaseId
)
2376 if (worker
->wi_tableoid
== relid
)
2379 found_concurrent_worker
= true;
2383 LWLockRelease(AutovacuumLock
);
2386 LWLockRelease(AutovacuumScheduleLock
);
2391 * Store the table's OID in shared memory before releasing the
2392 * schedule lock, so that other workers don't try to vacuum it
2393 * concurrently. (We claim it here so as not to hold
2394 * AutovacuumScheduleLock while rechecking the stats.)
2396 MyWorkerInfo
->wi_tableoid
= relid
;
2397 MyWorkerInfo
->wi_sharedrel
= isshared
;
2398 LWLockRelease(AutovacuumScheduleLock
);
2401 * Check whether pgstat data still says we need to vacuum this table.
2402 * It could have changed if something else processed the table while
2403 * we weren't looking. This doesn't entirely close the race condition,
2404 * but it is very small.
2406 MemoryContextSwitchTo(AutovacMemCxt
);
2407 tab
= table_recheck_autovac(relid
, table_toast_map
, pg_class_desc
,
2408 effective_multixact_freeze_max_age
);
2411 /* someone else vacuumed the table, or it went away */
2412 LWLockAcquire(AutovacuumScheduleLock
, LW_EXCLUSIVE
);
2413 MyWorkerInfo
->wi_tableoid
= InvalidOid
;
2414 MyWorkerInfo
->wi_sharedrel
= false;
2415 LWLockRelease(AutovacuumScheduleLock
);
2420 * Remember the prevailing values of the vacuum cost GUCs. We have to
2421 * restore these at the bottom of the loop, else we'll compute wrong
2422 * values in the next iteration of autovac_balance_cost().
2424 stdVacuumCostDelay
= VacuumCostDelay
;
2425 stdVacuumCostLimit
= VacuumCostLimit
;
2427 /* Must hold AutovacuumLock while mucking with cost balance info */
2428 LWLockAcquire(AutovacuumLock
, LW_EXCLUSIVE
);
2430 /* advertise my cost delay parameters for the balancing algorithm */
2431 MyWorkerInfo
->wi_dobalance
= tab
->at_dobalance
;
2432 MyWorkerInfo
->wi_cost_delay
= tab
->at_vacuum_cost_delay
;
2433 MyWorkerInfo
->wi_cost_limit
= tab
->at_vacuum_cost_limit
;
2434 MyWorkerInfo
->wi_cost_limit_base
= tab
->at_vacuum_cost_limit
;
2437 autovac_balance_cost();
2439 /* set the active cost parameters from the result of that */
2440 AutoVacuumUpdateDelay();
2443 LWLockRelease(AutovacuumLock
);
2445 /* clean up memory before each iteration */
2446 MemoryContextResetAndDeleteChildren(PortalContext
);
2449 * Save the relation name for a possible error message, to avoid a
2450 * catalog lookup in case of an error. If any of these return NULL,
2451 * then the relation has been dropped since last we checked; skip it.
2452 * Note: they must live in a long-lived memory context because we call
2453 * vacuum and analyze in different transactions.
2456 tab
->at_relname
= get_rel_name(tab
->at_relid
);
2457 tab
->at_nspname
= get_namespace_name(get_rel_namespace(tab
->at_relid
));
2458 tab
->at_datname
= get_database_name(MyDatabaseId
);
2459 if (!tab
->at_relname
|| !tab
->at_nspname
|| !tab
->at_datname
)
2463 * We will abort vacuuming the current table if something errors out,
2464 * and continue with the next one in schedule; in particular, this
2465 * happens if we are interrupted with SIGINT.
2469 /* Use PortalContext for any per-table allocations */
2470 MemoryContextSwitchTo(PortalContext
);
2473 autovacuum_do_vac_analyze(tab
, bstrategy
);
2476 * Clear a possible query-cancel signal, to avoid a late reaction
2477 * to an automatically-sent signal because of vacuuming the
2478 * current table (we're done with it, so it would make no sense to
2479 * cancel at this point.)
2481 QueryCancelPending
= false;
2486 * Abort the transaction, start a new one, and proceed with the
2487 * next table in our list.
2490 if (tab
->at_params
.options
& VACOPT_VACUUM
)
2491 errcontext("automatic vacuum of table \"%s.%s.%s\"",
2492 tab
->at_datname
, tab
->at_nspname
, tab
->at_relname
);
2494 errcontext("automatic analyze of table \"%s.%s.%s\"",
2495 tab
->at_datname
, tab
->at_nspname
, tab
->at_relname
);
2498 /* this resets ProcGlobal->statusFlags[i] too */
2499 AbortOutOfAnyTransaction();
2501 MemoryContextResetAndDeleteChildren(PortalContext
);
2503 /* restart our transaction for the following operations */
2504 StartTransactionCommand();
2505 RESUME_INTERRUPTS();
2509 /* Make sure we're back in AutovacMemCxt */
2510 MemoryContextSwitchTo(AutovacMemCxt
);
2514 /* ProcGlobal->statusFlags[i] are reset at the next end of xact */
2518 if (tab
->at_datname
!= NULL
)
2519 pfree(tab
->at_datname
);
2520 if (tab
->at_nspname
!= NULL
)
2521 pfree(tab
->at_nspname
);
2522 if (tab
->at_relname
!= NULL
)
2523 pfree(tab
->at_relname
);
2527 * Remove my info from shared memory. We could, but intentionally
2528 * don't, clear wi_cost_limit and friends --- this is on the
2529 * assumption that we probably have more to do with similar cost
2530 * settings, so we don't want to give up our share of I/O for a very
2531 * short interval and thereby thrash the global balance.
2533 LWLockAcquire(AutovacuumScheduleLock
, LW_EXCLUSIVE
);
2534 MyWorkerInfo
->wi_tableoid
= InvalidOid
;
2535 MyWorkerInfo
->wi_sharedrel
= false;
2536 LWLockRelease(AutovacuumScheduleLock
);
2538 /* restore vacuum cost GUCs for the next iteration */
2539 VacuumCostDelay
= stdVacuumCostDelay
;
2540 VacuumCostLimit
= stdVacuumCostLimit
;
2544 * Perform additional work items, as requested by backends.
2546 LWLockAcquire(AutovacuumLock
, LW_EXCLUSIVE
);
2547 for (i
= 0; i
< NUM_WORKITEMS
; i
++)
2549 AutoVacuumWorkItem
*workitem
= &AutoVacuumShmem
->av_workItems
[i
];
2551 if (!workitem
->avw_used
)
2553 if (workitem
->avw_active
)
2555 if (workitem
->avw_database
!= MyDatabaseId
)
2558 /* claim this one, and release lock while performing it */
2559 workitem
->avw_active
= true;
2560 LWLockRelease(AutovacuumLock
);
2562 perform_work_item(workitem
);
2565 * Check for config changes before acquiring lock for further jobs.
2567 CHECK_FOR_INTERRUPTS();
2568 if (ConfigReloadPending
)
2570 ConfigReloadPending
= false;
2571 ProcessConfigFile(PGC_SIGHUP
);
2574 LWLockAcquire(AutovacuumLock
, LW_EXCLUSIVE
);
2576 /* and mark it done */
2577 workitem
->avw_active
= false;
2578 workitem
->avw_used
= false;
2580 LWLockRelease(AutovacuumLock
);
2583 * We leak table_toast_map here (among other things), but since we're
2584 * going away soon, it's not a problem.
2588 * Update pg_database.datfrozenxid, and truncate pg_xact if possible. We
2589 * only need to do this once, not after each table.
2591 * Even if we didn't vacuum anything, it may still be important to do
2592 * this, because one indirect effect of vac_update_datfrozenxid() is to
2593 * update ShmemVariableCache->xidVacLimit. That might need to be done
2594 * even if we haven't vacuumed anything, because relations with older
2595 * relfrozenxid values or other databases with older datfrozenxid values
2596 * might have been dropped, allowing xidVacLimit to advance.
2598 * However, it's also important not to do this blindly in all cases,
2599 * because when autovacuum=off this will restart the autovacuum launcher.
2600 * If we're not careful, an infinite loop can result, where workers find
2601 * no work to do and restart the launcher, which starts another worker in
2602 * the same database that finds no work to do. To prevent that, we skip
2603 * this if (1) we found no work to do and (2) we skipped at least one
2604 * table due to concurrent autovacuum activity. In that case, the other
2605 * worker has already done it, or will do so when it finishes.
2607 if (did_vacuum
|| !found_concurrent_worker
)
2608 vac_update_datfrozenxid();
2610 /* Finally close out the last transaction. */
2611 CommitTransactionCommand();
2615 * Execute a previously registered work item.
2618 perform_work_item(AutoVacuumWorkItem
*workitem
)
2620 char *cur_datname
= NULL
;
2621 char *cur_nspname
= NULL
;
2622 char *cur_relname
= NULL
;
2625 * Note we do not store table info in MyWorkerInfo, since this is not
2630 * Save the relation name for a possible error message, to avoid a catalog
2631 * lookup in case of an error. If any of these return NULL, then the
2632 * relation has been dropped since last we checked; skip it.
2634 Assert(CurrentMemoryContext
== AutovacMemCxt
);
2636 cur_relname
= get_rel_name(workitem
->avw_relation
);
2637 cur_nspname
= get_namespace_name(get_rel_namespace(workitem
->avw_relation
));
2638 cur_datname
= get_database_name(MyDatabaseId
);
2639 if (!cur_relname
|| !cur_nspname
|| !cur_datname
)
2642 autovac_report_workitem(workitem
, cur_nspname
, cur_relname
);
2644 /* clean up memory before each work item */
2645 MemoryContextResetAndDeleteChildren(PortalContext
);
2648 * We will abort the current work item if something errors out, and
2649 * continue with the next one; in particular, this happens if we are
2650 * interrupted with SIGINT. Note that this means that the work item list
2655 /* Use PortalContext for any per-work-item allocations */
2656 MemoryContextSwitchTo(PortalContext
);
2659 switch (workitem
->avw_type
)
2661 case AVW_BRINSummarizeRange
:
2662 DirectFunctionCall2(brin_summarize_range
,
2663 ObjectIdGetDatum(workitem
->avw_relation
),
2664 Int64GetDatum((int64
) workitem
->avw_blockNumber
));
2667 elog(WARNING
, "unrecognized work item found: type %d",
2668 workitem
->avw_type
);
2673 * Clear a possible query-cancel signal, to avoid a late reaction to
2674 * an automatically-sent signal because of vacuuming the current table
2675 * (we're done with it, so it would make no sense to cancel at this
2678 QueryCancelPending
= false;
2683 * Abort the transaction, start a new one, and proceed with the next
2684 * table in our list.
2687 errcontext("processing work entry for relation \"%s.%s.%s\"",
2688 cur_datname
, cur_nspname
, cur_relname
);
2691 /* this resets ProcGlobal->statusFlags[i] too */
2692 AbortOutOfAnyTransaction();
2694 MemoryContextResetAndDeleteChildren(PortalContext
);
2696 /* restart our transaction for the following operations */
2697 StartTransactionCommand();
2698 RESUME_INTERRUPTS();
2702 /* Make sure we're back in AutovacMemCxt */
2703 MemoryContextSwitchTo(AutovacMemCxt
);
2705 /* We intentionally do not set did_vacuum here */
2718 * extract_autovac_opts
2720 * Given a relation's pg_class tuple, return the AutoVacOpts portion of
2721 * reloptions, if set; otherwise, return NULL.
2723 * Note: callers do not have a relation lock on the table at this point,
2724 * so the table could have been dropped, and its catalog rows gone, after
2725 * we acquired the pg_class row. If pg_class had a TOAST table, this would
2726 * be a risk; fortunately, it doesn't.
2728 static AutoVacOpts
*
2729 extract_autovac_opts(HeapTuple tup
, TupleDesc pg_class_desc
)
2734 Assert(((Form_pg_class
) GETSTRUCT(tup
))->relkind
== RELKIND_RELATION
||
2735 ((Form_pg_class
) GETSTRUCT(tup
))->relkind
== RELKIND_MATVIEW
||
2736 ((Form_pg_class
) GETSTRUCT(tup
))->relkind
== RELKIND_TOASTVALUE
);
2738 relopts
= extractRelOptions(tup
, pg_class_desc
, NULL
);
2739 if (relopts
== NULL
)
2742 av
= palloc(sizeof(AutoVacOpts
));
2743 memcpy(av
, &(((StdRdOptions
*) relopts
)->autovacuum
), sizeof(AutoVacOpts
));
2751 * table_recheck_autovac
2753 * Recheck whether a table still needs vacuum or analyze. Return value is a
2754 * valid autovac_table pointer if it does, NULL otherwise.
2756 * Note that the returned autovac_table does not have the name fields set.
2758 static autovac_table
*
2759 table_recheck_autovac(Oid relid
, HTAB
*table_toast_map
,
2760 TupleDesc pg_class_desc
,
2761 int effective_multixact_freeze_max_age
)
2763 Form_pg_class classForm
;
2767 autovac_table
*tab
= NULL
;
2769 AutoVacOpts
*avopts
;
2771 /* fetch the relation's relcache entry */
2772 classTup
= SearchSysCacheCopy1(RELOID
, ObjectIdGetDatum(relid
));
2773 if (!HeapTupleIsValid(classTup
))
2775 classForm
= (Form_pg_class
) GETSTRUCT(classTup
);
2778 * Get the applicable reloptions. If it is a TOAST table, try to get the
2779 * main table reloptions if the toast table itself doesn't have.
2781 avopts
= extract_autovac_opts(classTup
, pg_class_desc
);
2782 if (classForm
->relkind
== RELKIND_TOASTVALUE
&&
2783 avopts
== NULL
&& table_toast_map
!= NULL
)
2785 av_relation
*hentry
;
2788 hentry
= hash_search(table_toast_map
, &relid
, HASH_FIND
, &found
);
2789 if (found
&& hentry
->ar_hasrelopts
)
2790 avopts
= &hentry
->ar_reloptions
;
2793 recheck_relation_needs_vacanalyze(relid
, avopts
, classForm
,
2794 effective_multixact_freeze_max_age
,
2795 &dovacuum
, &doanalyze
, &wraparound
);
2797 /* OK, it needs something done */
2798 if (doanalyze
|| dovacuum
)
2801 int freeze_table_age
;
2802 int multixact_freeze_min_age
;
2803 int multixact_freeze_table_age
;
2805 double vac_cost_delay
;
2806 int log_min_duration
;
2809 * Calculate the vacuum cost parameters and the freeze ages. If there
2810 * are options set in pg_class.reloptions, use them; in the case of a
2811 * toast table, try the main table too. Otherwise use the GUC
2812 * defaults, autovacuum's own first and plain vacuum second.
2815 /* -1 in autovac setting means use plain vacuum_cost_delay */
2816 vac_cost_delay
= (avopts
&& avopts
->vacuum_cost_delay
>= 0)
2817 ? avopts
->vacuum_cost_delay
2818 : (autovacuum_vac_cost_delay
>= 0)
2819 ? autovacuum_vac_cost_delay
2822 /* 0 or -1 in autovac setting means use plain vacuum_cost_limit */
2823 vac_cost_limit
= (avopts
&& avopts
->vacuum_cost_limit
> 0)
2824 ? avopts
->vacuum_cost_limit
2825 : (autovacuum_vac_cost_limit
> 0)
2826 ? autovacuum_vac_cost_limit
2829 /* -1 in autovac setting means use log_autovacuum_min_duration */
2830 log_min_duration
= (avopts
&& avopts
->log_min_duration
>= 0)
2831 ? avopts
->log_min_duration
2832 : Log_autovacuum_min_duration
;
2834 /* these do not have autovacuum-specific settings */
2835 freeze_min_age
= (avopts
&& avopts
->freeze_min_age
>= 0)
2836 ? avopts
->freeze_min_age
2837 : default_freeze_min_age
;
2839 freeze_table_age
= (avopts
&& avopts
->freeze_table_age
>= 0)
2840 ? avopts
->freeze_table_age
2841 : default_freeze_table_age
;
2843 multixact_freeze_min_age
= (avopts
&&
2844 avopts
->multixact_freeze_min_age
>= 0)
2845 ? avopts
->multixact_freeze_min_age
2846 : default_multixact_freeze_min_age
;
2848 multixact_freeze_table_age
= (avopts
&&
2849 avopts
->multixact_freeze_table_age
>= 0)
2850 ? avopts
->multixact_freeze_table_age
2851 : default_multixact_freeze_table_age
;
2853 tab
= palloc(sizeof(autovac_table
));
2854 tab
->at_relid
= relid
;
2855 tab
->at_sharedrel
= classForm
->relisshared
;
2858 * Select VACUUM options. Note we don't say VACOPT_PROCESS_TOAST, so
2859 * that vacuum() skips toast relations. Also note we tell vacuum() to
2860 * skip vac_update_datfrozenxid(); we'll do that separately.
2862 tab
->at_params
.options
=
2863 (dovacuum
? (VACOPT_VACUUM
|
2864 VACOPT_PROCESS_MAIN
|
2865 VACOPT_SKIP_DATABASE_STATS
) : 0) |
2866 (doanalyze
? VACOPT_ANALYZE
: 0) |
2867 (!wraparound
? VACOPT_SKIP_LOCKED
: 0);
2870 * index_cleanup and truncate are unspecified at first in autovacuum.
2871 * They will be filled in with usable values using their reloptions
2872 * (or reloption defaults) later.
2874 tab
->at_params
.index_cleanup
= VACOPTVALUE_UNSPECIFIED
;
2875 tab
->at_params
.truncate
= VACOPTVALUE_UNSPECIFIED
;
2876 /* As of now, we don't support parallel vacuum for autovacuum */
2877 tab
->at_params
.nworkers
= -1;
2878 tab
->at_params
.freeze_min_age
= freeze_min_age
;
2879 tab
->at_params
.freeze_table_age
= freeze_table_age
;
2880 tab
->at_params
.multixact_freeze_min_age
= multixact_freeze_min_age
;
2881 tab
->at_params
.multixact_freeze_table_age
= multixact_freeze_table_age
;
2882 tab
->at_params
.is_wraparound
= wraparound
;
2883 tab
->at_params
.log_min_duration
= log_min_duration
;
2884 tab
->at_vacuum_cost_limit
= vac_cost_limit
;
2885 tab
->at_vacuum_cost_delay
= vac_cost_delay
;
2886 tab
->at_relname
= NULL
;
2887 tab
->at_nspname
= NULL
;
2888 tab
->at_datname
= NULL
;
2891 * If any of the cost delay parameters has been set individually for
2892 * this table, disable the balancing algorithm.
2895 !(avopts
&& (avopts
->vacuum_cost_limit
> 0 ||
2896 avopts
->vacuum_cost_delay
> 0));
2899 heap_freetuple(classTup
);
2904 * recheck_relation_needs_vacanalyze
2906 * Subroutine for table_recheck_autovac.
2908 * Fetch the pgstat of a relation and recheck whether a relation
2909 * needs to be vacuumed or analyzed.
2912 recheck_relation_needs_vacanalyze(Oid relid
,
2913 AutoVacOpts
*avopts
,
2914 Form_pg_class classForm
,
2915 int effective_multixact_freeze_max_age
,
2920 PgStat_StatTabEntry
*tabentry
;
2922 /* fetch the pgstat table entry */
2923 tabentry
= pgstat_fetch_stat_tabentry_ext(classForm
->relisshared
,
2926 relation_needs_vacanalyze(relid
, avopts
, classForm
, tabentry
,
2927 effective_multixact_freeze_max_age
,
2928 dovacuum
, doanalyze
, wraparound
);
2930 /* ignore ANALYZE for toast tables */
2931 if (classForm
->relkind
== RELKIND_TOASTVALUE
)
2936 * relation_needs_vacanalyze
2938 * Check whether a relation needs to be vacuumed or analyzed; return each into
2939 * "dovacuum" and "doanalyze", respectively. Also return whether the vacuum is
2940 * being forced because of Xid or multixact wraparound.
2942 * relopts is a pointer to the AutoVacOpts options (either for itself in the
2943 * case of a plain table, or for either itself or its parent table in the case
2944 * of a TOAST table), NULL if none; tabentry is the pgstats entry, which can be
2947 * A table needs to be vacuumed if the number of dead tuples exceeds a
2948 * threshold. This threshold is calculated as
2950 * threshold = vac_base_thresh + vac_scale_factor * reltuples
2952 * For analyze, the analysis done is that the number of tuples inserted,
2953 * deleted and updated since the last analyze exceeds a threshold calculated
2954 * in the same fashion as above. Note that the cumulative stats system stores
2955 * the number of tuples (both live and dead) that there were as of the last
2956 * analyze. This is asymmetric to the VACUUM case.
2958 * We also force vacuum if the table's relfrozenxid is more than freeze_max_age
2959 * transactions back, and if its relminmxid is more than
2960 * multixact_freeze_max_age multixacts back.
2962 * A table whose autovacuum_enabled option is false is
2963 * automatically skipped (unless we have to vacuum it due to freeze_max_age).
2964 * Thus autovacuum can be disabled for specific tables. Also, when the cumulative
2965 * stats system does not have data about a table, it will be skipped.
2967 * A table whose vac_base_thresh value is < 0 takes the base value from the
2968 * autovacuum_vacuum_threshold GUC variable. Similarly, a vac_scale_factor
2969 * value < 0 is substituted with the value of
2970 * autovacuum_vacuum_scale_factor GUC variable. Ditto for analyze.
2973 relation_needs_vacanalyze(Oid relid
,
2974 AutoVacOpts
*relopts
,
2975 Form_pg_class classForm
,
2976 PgStat_StatTabEntry
*tabentry
,
2977 int effective_multixact_freeze_max_age
,
2978 /* output params below */
2985 float4 reltuples
; /* pg_class.reltuples */
2987 /* constants from reloptions or GUC variables */
2988 int vac_base_thresh
,
2989 vac_ins_base_thresh
,
2991 float4 vac_scale_factor
,
2992 vac_ins_scale_factor
,
2995 /* thresholds calculated from above constants */
3000 /* number of vacuum (resp. analyze) tuples at this time */
3005 /* freeze parameters */
3007 int multixact_freeze_max_age
;
3008 TransactionId xidForceLimit
;
3009 MultiXactId multiForceLimit
;
3011 Assert(classForm
!= NULL
);
3012 Assert(OidIsValid(relid
));
3015 * Determine vacuum/analyze equation parameters. We have two possible
3016 * sources: the passed reloptions (which could be a main table or a toast
3017 * table), or the autovacuum GUC variables.
3020 /* -1 in autovac setting means use plain vacuum_scale_factor */
3021 vac_scale_factor
= (relopts
&& relopts
->vacuum_scale_factor
>= 0)
3022 ? relopts
->vacuum_scale_factor
3023 : autovacuum_vac_scale
;
3025 vac_base_thresh
= (relopts
&& relopts
->vacuum_threshold
>= 0)
3026 ? relopts
->vacuum_threshold
3027 : autovacuum_vac_thresh
;
3029 vac_ins_scale_factor
= (relopts
&& relopts
->vacuum_ins_scale_factor
>= 0)
3030 ? relopts
->vacuum_ins_scale_factor
3031 : autovacuum_vac_ins_scale
;
3033 /* -1 is used to disable insert vacuums */
3034 vac_ins_base_thresh
= (relopts
&& relopts
->vacuum_ins_threshold
>= -1)
3035 ? relopts
->vacuum_ins_threshold
3036 : autovacuum_vac_ins_thresh
;
3038 anl_scale_factor
= (relopts
&& relopts
->analyze_scale_factor
>= 0)
3039 ? relopts
->analyze_scale_factor
3040 : autovacuum_anl_scale
;
3042 anl_base_thresh
= (relopts
&& relopts
->analyze_threshold
>= 0)
3043 ? relopts
->analyze_threshold
3044 : autovacuum_anl_thresh
;
3046 freeze_max_age
= (relopts
&& relopts
->freeze_max_age
>= 0)
3047 ? Min(relopts
->freeze_max_age
, autovacuum_freeze_max_age
)
3048 : autovacuum_freeze_max_age
;
3050 multixact_freeze_max_age
= (relopts
&& relopts
->multixact_freeze_max_age
>= 0)
3051 ? Min(relopts
->multixact_freeze_max_age
, effective_multixact_freeze_max_age
)
3052 : effective_multixact_freeze_max_age
;
3054 av_enabled
= (relopts
? relopts
->enabled
: true);
3056 /* Force vacuum if table is at risk of wraparound */
3057 xidForceLimit
= recentXid
- freeze_max_age
;
3058 if (xidForceLimit
< FirstNormalTransactionId
)
3059 xidForceLimit
-= FirstNormalTransactionId
;
3060 force_vacuum
= (TransactionIdIsNormal(classForm
->relfrozenxid
) &&
3061 TransactionIdPrecedes(classForm
->relfrozenxid
,
3065 multiForceLimit
= recentMulti
- multixact_freeze_max_age
;
3066 if (multiForceLimit
< FirstMultiXactId
)
3067 multiForceLimit
-= FirstMultiXactId
;
3068 force_vacuum
= MultiXactIdIsValid(classForm
->relminmxid
) &&
3069 MultiXactIdPrecedes(classForm
->relminmxid
, multiForceLimit
);
3071 *wraparound
= force_vacuum
;
3073 /* User disabled it in pg_class.reloptions? (But ignore if at risk) */
3074 if (!av_enabled
&& !force_vacuum
)
3082 * If we found stats for the table, and autovacuum is currently enabled,
3083 * make a threshold-based decision whether to vacuum and/or analyze. If
3084 * autovacuum is currently disabled, we must be here for anti-wraparound
3085 * vacuuming only, so don't vacuum (or analyze) anything that's not being
3088 if (PointerIsValid(tabentry
) && AutoVacuumingActive())
3090 reltuples
= classForm
->reltuples
;
3091 vactuples
= tabentry
->dead_tuples
;
3092 instuples
= tabentry
->ins_since_vacuum
;
3093 anltuples
= tabentry
->mod_since_analyze
;
3095 /* If the table hasn't yet been vacuumed, take reltuples as zero */
3099 vacthresh
= (float4
) vac_base_thresh
+ vac_scale_factor
* reltuples
;
3100 vacinsthresh
= (float4
) vac_ins_base_thresh
+ vac_ins_scale_factor
* reltuples
;
3101 anlthresh
= (float4
) anl_base_thresh
+ anl_scale_factor
* reltuples
;
3104 * Note that we don't need to take special consideration for stat
3105 * reset, because if that happens, the last vacuum and analyze counts
3106 * will be reset too.
3108 if (vac_ins_base_thresh
>= 0)
3109 elog(DEBUG3
, "%s: vac: %.0f (threshold %.0f), ins: %.0f (threshold %.0f), anl: %.0f (threshold %.0f)",
3110 NameStr(classForm
->relname
),
3111 vactuples
, vacthresh
, instuples
, vacinsthresh
, anltuples
, anlthresh
);
3113 elog(DEBUG3
, "%s: vac: %.0f (threshold %.0f), ins: (disabled), anl: %.0f (threshold %.0f)",
3114 NameStr(classForm
->relname
),
3115 vactuples
, vacthresh
, anltuples
, anlthresh
);
3117 /* Determine if this table needs vacuum or analyze. */
3118 *dovacuum
= force_vacuum
|| (vactuples
> vacthresh
) ||
3119 (vac_ins_base_thresh
>= 0 && instuples
> vacinsthresh
);
3120 *doanalyze
= (anltuples
> anlthresh
);
3125 * Skip a table not found in stat hash, unless we have to force vacuum
3126 * for anti-wrap purposes. If it's not acted upon, there's no need to
3129 *dovacuum
= force_vacuum
;
3133 /* ANALYZE refuses to work with pg_statistic */
3134 if (relid
== StatisticRelationId
)
3139 * autovacuum_do_vac_analyze
3140 * Vacuum and/or analyze the specified table
3143 autovacuum_do_vac_analyze(autovac_table
*tab
, BufferAccessStrategy bstrategy
)
3146 VacuumRelation
*rel
;
3149 /* Let pgstat know what we're doing */
3150 autovac_report_activity(tab
);
3152 /* Set up one VacuumRelation target, identified by OID, for vacuum() */
3153 rangevar
= makeRangeVar(tab
->at_nspname
, tab
->at_relname
, -1);
3154 rel
= makeVacuumRelation(rangevar
, tab
->at_relid
, NIL
);
3155 rel_list
= list_make1(rel
);
3157 vacuum(rel_list
, &tab
->at_params
, bstrategy
, true);
3161 * autovac_report_activity
3162 * Report to pgstat what autovacuum is doing
3164 * We send a SQL string corresponding to what the user would see if the
3165 * equivalent command was to be issued manually.
3167 * Note we assume that we are going to report the next command as soon as we're
3168 * done with the current one, and exit right after the last one, so we don't
3169 * bother to report "<IDLE>" or some such.
3172 autovac_report_activity(autovac_table
*tab
)
3174 #define MAX_AUTOVAC_ACTIV_LEN (NAMEDATALEN * 2 + 56)
3175 char activity
[MAX_AUTOVAC_ACTIV_LEN
];
3178 /* Report the command and possible options */
3179 if (tab
->at_params
.options
& VACOPT_VACUUM
)
3180 snprintf(activity
, MAX_AUTOVAC_ACTIV_LEN
,
3181 "autovacuum: VACUUM%s",
3182 tab
->at_params
.options
& VACOPT_ANALYZE
? " ANALYZE" : "");
3184 snprintf(activity
, MAX_AUTOVAC_ACTIV_LEN
,
3185 "autovacuum: ANALYZE");
3188 * Report the qualified name of the relation.
3190 len
= strlen(activity
);
3192 snprintf(activity
+ len
, MAX_AUTOVAC_ACTIV_LEN
- len
,
3193 " %s.%s%s", tab
->at_nspname
, tab
->at_relname
,
3194 tab
->at_params
.is_wraparound
? " (to prevent wraparound)" : "");
3196 /* Set statement_timestamp() to current time for pg_stat_activity */
3197 SetCurrentStatementStartTimestamp();
3199 pgstat_report_activity(STATE_RUNNING
, activity
);
3203 * autovac_report_workitem
3204 * Report to pgstat that autovacuum is processing a work item
3207 autovac_report_workitem(AutoVacuumWorkItem
*workitem
,
3208 const char *nspname
, const char *relname
)
3210 char activity
[MAX_AUTOVAC_ACTIV_LEN
+ 12 + 2];
3214 switch (workitem
->avw_type
)
3216 case AVW_BRINSummarizeRange
:
3217 snprintf(activity
, MAX_AUTOVAC_ACTIV_LEN
,
3218 "autovacuum: BRIN summarize");
3223 * Report the qualified name of the relation, and the block number if any
3225 len
= strlen(activity
);
3227 if (BlockNumberIsValid(workitem
->avw_blockNumber
))
3228 snprintf(blk
, sizeof(blk
), " %u", workitem
->avw_blockNumber
);
3232 snprintf(activity
+ len
, MAX_AUTOVAC_ACTIV_LEN
- len
,
3233 " %s.%s%s", nspname
, relname
, blk
);
3235 /* Set statement_timestamp() to current time for pg_stat_activity */
3236 SetCurrentStatementStartTimestamp();
3238 pgstat_report_activity(STATE_RUNNING
, activity
);
3242 * AutoVacuumingActive
3243 * Check GUC vars and report whether the autovacuum process should be
3247 AutoVacuumingActive(void)
3249 if (!autovacuum_start_daemon
|| !pgstat_track_counts
)
3255 * Request one work item to the next autovacuum run processing our database.
3256 * Return false if the request can't be recorded.
3259 AutoVacuumRequestWork(AutoVacuumWorkItemType type
, Oid relationId
,
3263 bool result
= false;
3265 LWLockAcquire(AutovacuumLock
, LW_EXCLUSIVE
);
3268 * Locate an unused work item and fill it with the given data.
3270 for (i
= 0; i
< NUM_WORKITEMS
; i
++)
3272 AutoVacuumWorkItem
*workitem
= &AutoVacuumShmem
->av_workItems
[i
];
3274 if (workitem
->avw_used
)
3277 workitem
->avw_used
= true;
3278 workitem
->avw_active
= false;
3279 workitem
->avw_type
= type
;
3280 workitem
->avw_database
= MyDatabaseId
;
3281 workitem
->avw_relation
= relationId
;
3282 workitem
->avw_blockNumber
= blkno
;
3289 LWLockRelease(AutovacuumLock
);
3296 * This is called at postmaster initialization.
3298 * All we do here is annoy the user if he got it wrong.
3303 if (autovacuum_start_daemon
&& !pgstat_track_counts
)
3305 (errmsg("autovacuum not started because of misconfiguration"),
3306 errhint("Enable the \"track_counts\" option.")));
3310 * IsAutoVacuum functions
3311 * Return whether this is either a launcher autovacuum process or a worker
3315 IsAutoVacuumLauncherProcess(void)
3317 return am_autovacuum_launcher
;
3321 IsAutoVacuumWorkerProcess(void)
3323 return am_autovacuum_worker
;
3328 * AutoVacuumShmemSize
3329 * Compute space needed for autovacuum-related shared memory
3332 AutoVacuumShmemSize(void)
3337 * Need the fixed struct and the array of WorkerInfoData.
3339 size
= sizeof(AutoVacuumShmemStruct
);
3340 size
= MAXALIGN(size
);
3341 size
= add_size(size
, mul_size(autovacuum_max_workers
,
3342 sizeof(WorkerInfoData
)));
3347 * AutoVacuumShmemInit
3348 * Allocate and initialize autovacuum-related shared memory
3351 AutoVacuumShmemInit(void)
3355 AutoVacuumShmem
= (AutoVacuumShmemStruct
*)
3356 ShmemInitStruct("AutoVacuum Data",
3357 AutoVacuumShmemSize(),
3360 if (!IsUnderPostmaster
)
3367 AutoVacuumShmem
->av_launcherpid
= 0;
3368 dlist_init(&AutoVacuumShmem
->av_freeWorkers
);
3369 dlist_init(&AutoVacuumShmem
->av_runningWorkers
);
3370 AutoVacuumShmem
->av_startingWorker
= NULL
;
3371 memset(AutoVacuumShmem
->av_workItems
, 0,
3372 sizeof(AutoVacuumWorkItem
) * NUM_WORKITEMS
);
3374 worker
= (WorkerInfo
) ((char *) AutoVacuumShmem
+
3375 MAXALIGN(sizeof(AutoVacuumShmemStruct
)));
3377 /* initialize the WorkerInfo free list */
3378 for (i
= 0; i
< autovacuum_max_workers
; i
++)
3379 dlist_push_head(&AutoVacuumShmem
->av_freeWorkers
,
3380 &worker
[i
].wi_links
);
3387 * GUC check_hook for autovacuum_work_mem
3390 check_autovacuum_work_mem(int *newval
, void **extra
, GucSource source
)
3393 * -1 indicates fallback.
3395 * If we haven't yet changed the boot_val default of -1, just let it be.
3396 * Autovacuum will look to maintenance_work_mem instead.
3402 * We clamp manually-set values to at least 1MB. Since
3403 * maintenance_work_mem is always set to at least this value, do the same